Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust
310 KB
Raw Blame

Changelog: 2.x releases (archived)

The 2.x release notes, moved out of the root CHANGELOG.md when the next major line began. They are frozen like every released section.

2.6.7 - 2026-10-13

Deprecated

  • The 1.x compatibility-token commands user add, user expire, user revive and user rotate-token, and their admin endpoints, will be removed in 3.0.0. Tokens they issued keep working after the upgrade as ordinary API keys; manage them with api-key list|rotate|revoke, and add people with user add-human.

Fixed

  • Fixed the companion relay refusing to open its queue on Windows when another relay process was committing to the same queue. SQLite deletes its rollback journal at each commit, and Windows answers "access denied" for a file that is being deleted, so the queue-file guard failed with inspect relay.sqlite-journal: Access is denied. The guard now retries that answer for up to about 100 ms; a file that stays unreadable is still refused.
  • Fixed ai-memory hook running inside a helper container on Docker-wrapper installs (Linux/macOS). A hook config or plugin that called the wrapper paid about 280 ms per event, over the 200 ms hook budget, and its default 127.0.0.1 server URL pointed at the container instead of the host server. The wrapper now runs hook and hook-drain through the same checksum-verified native host client as mcp-bridge and install-hooks. A hook only reuses the installed client, with no update check or download on the hook path. When no client is installed yet, the wrapper says to run ai-memory install-hooks once, skips the event and exits 0, so the agent is never blocked.

2.6.6 - 2026-10-13

Fixed

  • Fixed ai-memory list-projects failing with 404 Not Found against a server started without --enable-web or --enable-api, the routes that mount GET /api/v1/projects. On a 404 it now reads the always-mounted GET /admin/projects and applies --workspace itself; ai-memory show lists projects the same way instead of falling back to local scan results. If both calls fail, the error names both routes and the likely causes. (#1236)
  • Fixed consolidation sending the model a prompt with no evidence when consolidation.max_input_tokens sat at the validated floor. The floor was checked against the raw token count, but the prompt builders work on a smaller character budget (tokens × a chars-per-token estimate × input_token_safety_margin, default 0.8). At the old 6,500 floor the batch prompt's fixed part alone overflowed that budget, so every observation became an omission marker. The floor is now 8,350 tokens at the default margin, and a lower margin needs a proportionally larger budget. A project's existing page titles are now optional prompt context, listed only as far as they fit, so a project with many pages no longer crowds observations out at a small budget. An install configured between 6,500 and 8,350 tokens now fails at startup with a message naming both numbers. (#1234, #1242)
  • Fixed PreCompact and PostCompaction waiting for the LLM provider inside /hook and /hook/batch. That 12-135 s call outlived the client's drain timeout, left the ingest key incomplete and restarted on every redelivery. The request now writes the rule-based sessions/<id>.md checkpoint and returns; with a provider, the LLM rewrite is queued for the bounded worker SessionEnd already uses, which now starts whenever a provider is configured and runs each job as the session's recorded owner. A resumed session's checkpoint queues its rewrite too, and a retried checkpoint never puts the rule-based page back over a finished rewrite. Two side effects: SessionEnd jobs queued while AI_MEMORY_CONSOLIDATE_ON_SESSION_END was on now run even if the flag has since been turned off, and without the flag a session that ends while its checkpoint rewrite runs keeps that rewrite as the latest page, with the SessionEnd summary as the superseded version. (#1229, #1243)
  • Replaced the unmaintained YAML library (serde_yaml and its unsafe-libyaml parser, RUSTSEC-2024-0449) with serde-saphyr. The server reads existing page frontmatter to exactly the same values as before. One formatting change: when a page is next saved, a frontmatter value that needs quoting (a title containing : , for example) is written in double quotes instead of single quotes. ai-memory-importer reads three rare hand-written forms differently: a numeric-looking title or tag such as 007 is no longer used, a !!timestamp-tagged title fails that page's import, and a custom-tagged title (!foo x) imports as its plain text.
  • Fixed the web search page printing "50 results" with no hint when more pages matched. It now says it shows the first 50 results and suggests refining the search. (#1239)
  • Fixed the web page header and the /api/v1 page response reporting the date of the latest version as "Created", so a page rewritten many times looked new. They now report when the oldest retained version was written. "Updated" is unchanged. (#1238)
  • Fixed keyboard and screen reader support in the web UI. Every page now has a "Skip to content" link, the search fields have labels, and breadcrumbs are a labelled landmark that marks the current page. Footer text, page paths and relative dates now meet AA contrast in both themes. (#1237)

2.6.5 - 2026-10-12

Fixed

  • Fixed the detached hook drainer recreating a deleted data directory. A drainer started just before ai-memory uninstall --purge-data (or one whose data directory was removed some other way) brought back an empty hook-spool/ with its lock file. The drainer now exits when the data directory is gone. A spool directory the drainer creates is also private (0700) now, as one created by a queued event always was; before, it took the default permissions.
  • Fixed the profile harvest writing a fresh copy of a stack preference on every pass. A statement classified stack (for example "Use Rust by default for everything else.") lives in profile/stack/ beside the per-language stack entries, and the step that finds an entry's existing page skipped that whole directory, so each harvest saw no page and wrote another -2, -3, … copy, each one shown in the session-start profile digest. Only the per-language pages (profile/stack/rust.md and the like) are skipped now. Copies already written stay until they are deleted.

2.6.4 - 2026-10-12

Fixed

  • Fixed Gemini CLI's run_shell_command and read_many_files tools being classified as unknown, so [capture] ignore_paths never checked them and a call reading an ignored file still had its output written to the local hook spool and sent to the server, which then discarded it. run_shell_command is now a shell tool, read like Bash and resolved from its dir_path, and read_many_files is a file tool: the call is dropped when any of its include globs names or could expand into an ignored path, and glob syntax the matcher cannot evaluate makes it metadata-only. Gemini's list_directory and the legacy search_file_content name are now search tools, dropped under an active policy like grep_search and glob. A shell command whose ** glob could recurse into an ignored directory (zsh, or bash with globstar) is now dropped too. A read_many_files entry naming a directory above an ignored path (docs, .), which Gemini reads recursively, is dropped as well. GitHub Copilot CLI's view, create and str_replace_editor are now file tools, checked against ignore_paths like Read and Write. Under an active or invalid policy, tools that read back output of an earlier command (Claude Code BashOutput/TaskOutput, Codex write_stdin, Hermes and OpenClaw process, Copilot CLI read_bash and write_bash) and shell commands using brace expansion or a bracket class are now metadata-only, since the matcher cannot prove they miss an ignored path; upgrade the server before the clients, because an older server drops those events. A script quoted inside a string-form shell command (an sh -c argument, a command substitution in double quotes) is now read like the rest of the command line, the path in a git object name (REV:path) is checked, Antigravity's run_command resolves its arguments from its Cwd, and ANSI-C strings and backslash-escaped letters are read the way the shell reads them. Inline interpreter code (python3 -c, node -e and similar) and further glob syntax the matcher does not evaluate (zsh glob groups, extglobs, encoded escapes) are now metadata-only under an active policy. Line continuations, parameter defaults, values glued to short flags, file: URLs, ~+, PowerShell and cmd escape characters, and shell or interpreter options placed before the script or code are now read too, and more script runners and interpreters are recognized. A path under a directory only the running shell knows (~-, ~name) makes the call metadata-only, and on macOS paths match without regard to case, as its default filesystem does, including non-ASCII letters and decomposed accents. A checkout's ignore_paths now also hold for a file, search/list or shell tool call made from another checkout or from an unmarked directory that reaches one of its files, including through glob syntax in a directory name, a search's path or file filter, or the files an apply_patch names. Glob directories outside the event's own checkout are followed through the directories on disk within fixed bounds of count and time; past them, or for glob syntax that cannot be followed, the call is metadata-only, and so is a file tool whose target cannot be read. A glob inside the event's own checkout is not walked, so ordinary monorepo commands over many packages (cat packages/*/package.json, du -sh packages/*) keep their output with or without a marker, and neither are a large brace expansion that cannot leave it (touch test/file{1..200}.txt) or the checkout's own subtree under a ** that starts above it. Each directory's marker is looked up once, so many paths in a deep tree stay within the hook's time budget, and a dotfiles .git in the home directory no longer counts as every project's own checkout. A .. after a **, a dot glob that older bash expands to the parent directory (POSIX classes and extglobs included), a brace expansion, a zsh glob group, an ANSI-C escape (also inside a double-quoted bash -c or eval script), an indented apply_patch header and Hermes search_files' file_glob filter can no longer hide which checkout a call reaches. A broken marker makes a search metadata-only, as it does a file or shell call. Reading back a file a harness saved an earlier tool's output to (Claude Code's saved large results and background task output, OpenCode's and Pi's truncated output files), or a glob that could match one, is metadata-only under an active or invalid policy, like BashOutput; an ordinary log glob (tail -n 50 log/*.log, cat /tmp/*.log) is not taken for one. A file: URL's query or fragment is no longer read as part of its path, a parameter default nested deeper than the matcher reads is metadata-only, and a backslash escape inside a default is read. A quoted ~ (git commit -m "fix ~approx") is no longer treated as a home or user directory, so such commands are kept again. In a Windows session, Git Bash, MSYS2 and Cygwin spellings (/c/..., /cygdrive/c/...), a rooted path without a drive, the \\?\, \\.\ and \\localhost\C$ prefixes, an alternate data stream suffix and trailing dots or spaces are read as the file they open, so they no longer slip past a pattern. The cwd, the marker's directory and an absolute pattern get the same reading, so a session started from \\?\C:\repo or \\localhost\C$\repo matches its relative patterns again. A bare or double-quoted \\server\share\... or \\?\C:\... argument keeps both backslashes, as PowerShell and cmd read it, and /a/x is also read from the root of the cwd's drive, as a native tool reads it. The generated TypeScript marker reader no longer takes a line inside a multi-line array (["docs"]) for a table header. Hermes execute_code is metadata-only under an active policy, like python3 -c, and a shell command longer than 1 MiB is metadata-only instead of being split. The generated TypeScript integrations now read a dotted capture.ignore_paths key and quoted [capture] headers, reject the shapes the native hook rejects (an inline capture = { ... } table among them), recognize the same file tools (read_file, patch, fs_read, ...), and stop reading a directory at the walk's entry bound. Checking a path of thousands of wildcard segments for a saved output file no longer takes seconds. In a Windows session, a shell argument or workdir that is drive-relative (C:docs\x) or a device path naming no drive (\\?\GLOBALROOT\...) now makes the command metadata-only instead of being skipped, a PowerShell FileSystem:: provider prefix is read as the path after it, and a Windows marker holding a POSIX absolute pattern (/c/repo/docs/**), which matched nothing, is now invalid and fails closed. This applies to the native hook, the server, and the generated TypeScript integrations. (GHSA-gcpc-88hm-3xqr)

  • Fixed ai-memory backfill, including the automatic first-boot import, and the ai-memory run workstream ledger not applying [capture] ignore_paths to the harness transcripts they import, so the arguments and output of a tool call on an ignored file could be stored even though live hooks drop them. Both now apply the same decision as the live hook to every imported tool call and its result: an excluded call is not imported at all, one the policy cannot prove safe keeps only its tool family, and prompts and assistant messages are imported as before, and under allowlist capture mode a checkout without a marker imports nothing. Results recorded without an id linking them to their call take the strictest decision of their turn. A call the transcript names differently from the hook (Codex's shell tool) is judged under the hook's name, each event is judged from the working directory its own record names, and a call or event the live hook would deliver to another checkout is imported only when that checkout would capture it and sends to the same server and project; events recorded in, or calls reaching into, a checkout of another project are left out instead of being filed under the importing project, and so are events from a different subdirectory of the same repository, which the server files under a project of their own. The output of a Claude Code ! shell command, which live capture never receives, is no longer imported. Claude Code's background-task notifications (a monitored command's output, a finished background command or agent) are no longer imported as prompts, whatever the case or spacing of their opening tag, and its compaction summary is imported as a compaction record instead of a prompt. A tool ai-memory does not recognize, such as an MCP tool, is imported as a record that it ran without its arguments, output or name, which is all live capture stores for it, and so is a result whose call is not in the transcript. (GHSA-gcpc-88hm-3xqr)

  • Fixed a web page open in the operator's browser being able to act on an unauthenticated server (the loopback default, no token), such as claiming a project's pending handoff or starting admin jobs. With auth off, the machine and admin routes now refuse browser requests that come from another site or that are not script calls from a page, such as links and images; hooks, the CLI, MCP clients and the built-in web UI are unaffected, and servers with a bearer token are unchanged. Images in wiki pages rendered by the web UI are also limited to relative paths that stay inside the UI. When an unauthenticated server is reached through a non-loopback host name in its allowlist (a LAN address or name, or host.docker.internal), where browsers do not send the request metadata the check relies on, the session-start handoff fetch now also requires an X-AI-Memory-Client header that every ai-memory hook, CLI command and generated integration sends; the same holds for a request that arrives through a forwarding proxy or CDN (any of the common forwarding headers, including CF-Connecting-IP, True-Client-IP and X-Forwarded-Proto), even one that rewrites the host to the loopback address. Hooks on such a server must be re-staged with ai-memory install-hooks (or ai-memory upgrade); loopback installs need nothing. (GHSA-6mc4-46pr-w82w)

  • Fixed POST /auth/login and POST /auth/recovery accepting request bodies that are not declared as JSON, which a page on another site could use to sign the operator's browser in or to try passwords when human sign-in is enabled. Both routes now answer 415 unless the body is application/json; the built-in login page and the documented curl recovery already send that. (GHSA-6mc4-46pr-w82w)

  • Fixed the web UI allowing other sites to frame it. With auth off, a page that framed /web/pending could trick a click on Approve or Reject, since the framed page's own requests are same-origin. Every web page and API response now sends X-Frame-Options: DENY and Content-Security-Policy: frame-ancestors 'none'. (GHSA-6mc4-46pr-w82w)

  • Fixed the cross-project profile harvester treating pages the auto-improve scheduler auto-approved as curated human input: model output nobody reviewed fed the profile like a hand-written rule. A page stamped from an auto-improve proposal is now skipped when the proposal's decider is the scheduler's auto-approve actor; a page a human approved still counts, and a page whose proposal id resolves to no row (legacy or imported) stays curated. (#1211)

  • Fixed the zero-LLM session page staying silent about abrupt endings. The rule-based synthesis now adds a deterministic terminal-state line reusing the automatic handoff's heuristics: "Session ended without a normal stop; the final prompt was never answered." when the log ends on an unanswered question, the equivalent line when file activity was never followed by a stop, and a mid-task-exit line when a stop was recorded but no session end arrived. A session with a normal stop and end gets no line. (#1221)

  • Fixed consolidation dropping the tail of a session from the compiled page. The single and batch prompts now carry explicit rules: say so when the observation log ends on an unanswered prompt or without a normal stop, and record in-flight work (background jobs, watchers, timers) as pending state; when a later observation corrects an earlier "done" milestone, state only the final state with the superseded claim as history at most; preserve the last reported metrics of in-flight work (sizes, rates, counts, ETAs) verbatim; carry operator "run X when Y" instructions verbatim, with their preconditions, into a Next steps / Pending section; and, for sessions that operated on external systems, emit an "Access and verification" section — hosts, the per-host auth mechanism (never secrets or tokens themselves), job-state marker locations, staged-script paths — grounded only in what the observations show. (#1221)

  • Softened the consolidation prompts' wikilink quota that pushed models to invent links. The single, batch and bootstrap prompts told the model "2-5 well-chosen links per page beat exhaustive linking; zero links should be rare", which pressured it to add links just to hit the number. They now state that zero links is fine when nothing in the input genuinely relates, require a genuine reason visible in the input for every link, and forbid padding a page with links to hit a number; the #1207 restriction on cross-scope links whose target is not visible in the input is unchanged. The validated consolidation.max_input_tokens floor moved from 6,000 to 6,500 tokens so the larger prompts still leave room for observation evidence at the minimum. (#1219)

  • Fixed consolidation still writing invented [[_global:…]] / [[project:…]] wikilinks when a model ignored the #1207 prompt wording. The single-page, batch and bootstrap apply paths now resolve every cross-scope wikilink target against the store with no-create lookups before the write and rewrite an unresolvable one to its plain label text (an unresolvable relations: frontmatter entry is dropped), so a missing target can no longer mint the dangling cross-project edge that ai-memory lint reports as broken_link. Resolved cross-scope links survive byte-identical, and same-project pending links stay untouched by design. A link into a restricted project is always de-linked, so the result never reveals which pages exist there. (#1218)

  • Fixed consolidation inventing [[_global:…]] and [[project:…]] wikilinks to pages that never existed, which ai-memory lint then reported as broken_link. The single, batch and bootstrap prompts told the model that a missing link target was acceptable and gave it no list of cross-scope pages, so it invented targets to meet the 2-5 link guidance. They now allow a project: or _global: link only when the exact wiki page path appears in the input, which reduces but does not eliminate such links; a missing same-project target stays a pending link. (#1207)

  • Fixed the web UI answering HTTP 500 when a signed-in viewer opened a project or page by a name that two projects answer to. The page and project routes now return a 400 that names the clash, matching the API and MCP surfaces. (#1204)

  • Fixed the Jev choice reranker adapter silently fabricating scores from a malformed judge response: docs/examples/jev-reranker-adapter/jev_rerank_shim_choice.py defaulted missing or misnamed probabilities entries to 0.0, so an HTTP 200 with a broken payload produced a well-formed but meaningless ranking the server could not tell apart from a real judgement. The shim now validates the response (key set matches the candidate list, no duplicate JSON keys, numeric non-boolean values in [0, 1], NaN/Inf rejected, distribution sum near 1) and answers HTTP 500 on any violation, which is the failure mode ai-memory already degrades safely on. Includes stdlib unittest coverage in test_jev_rerank_shim_choice.py. (#1198)

  • Fixed Hermes Agent tool events being dropped under capture_mode = allowlist. Hermes reports its own process directory (usually ~) as the payload cwd and points each tool at a repository with an absolute workdir or path, so the marker lookup never found the repository's .ai-memory.toml. The native ai-memory hook now also routes a shell command by its absolute workdir and a search/list tool by its absolute path, alongside the file-tool targets it already followed, when that location proves one other repository or marker; relative locations still keep the payload cwd. A reroute is now refused when the payload cwd's own [capture] ignore_paths would drop the event, so an exclusion that names a path outside its checkout, or a source file read by a command whose workdir points at another project, is no longer judged only by the destination's policy. When the payload cwd's own capture policy is invalid, an event aimed at another repository is no longer rerouted; it stays with that source repository as metadata only. (#1199)

  • scope_refusal (the hooks-side mirror of the web surfaces refusal_response mapped in #1204) now answers 404 Not Found for typed not-found ScopeResolutionError variants and 400 Bad Request for typed bad-request variants. Previously every refusal that was not Forbidden fell through to 500 Internal Server Error, hiding a missing-id mistake from the caller and making it indistinguishable from a real fault. The four-branch shape (forbidden / not_found / bad_request / fallback) matches the web mapping from #1204; the typed is_* predicates the store already exposes carry the classification, so the fix is purely a branch reshuffle with no new predicate. Four unit tests in workstream.rs::tests cover the unknown-id (404), unknown-workspace (404), ambiguous-name (400), and forbidden (403, regression guard) cases. (#1209)

  • POST /admin/pending-writes/{id}/approve and /reject answered 500 Internal Server Error for an id that did not exist or belonged to a different workspace, because the routes skipped the lookup and let the wiki method surface a generic error. They now share the same pending_detail helper the /detail and /diff routes already used, so all four routes agree on 400 for an unparsable id, 404 for an unknown or foreign-scope id, and 500 only for unexpected internal errors. New tests in admin::tests cover the unknown, foreign-scope, and unparsable cases for both approve and reject. (#1205)

  • Fixed Kimi Code sessions resumed after SessionEnd never receiving an open handoff. Kimi Code reuses its native session id when a session is resumed, so the receiver row already had ended_at set when the resumed session asked for its handoff, and accept_handoff refused the claim with "an ended session cannot accept a handoff" — an error the /handoff route swallows as an empty 200, leaving the handoff open forever. AgentKind::reuses_session_id_after_end() now covers Kimi Code alongside Grok, so accepting reopens the ended receiver row. Every surrounding guard is unchanged: an ended session of any other agent is still final, a receiver in another project is still refused, and the claim-once guard still stops the resumed session from taking a second baton. Covered by an adversarial integration test in multi_session.rs and a unit test on the predicate itself.

  • Fixed agents filing cross-project reports as wiki pages. Asked to "send a report to the ai-memory inbox", an agent wrote pages into the target project with memory_write_page instead of calling memory_message_send, and the report never existed as a message. The always-visible surfaces now carry the channel contrast in both directions: memory_message_send's tool description and its MEMORY_INSTRUCTIONS bullet name the trigger phrasing ("send a report", "notify", "leave a message for") and state it is the only tool that lands in a project's inbox; memory_write_page's description and bullet say a page is not a message and redirect messaging to memory_message_send. The routing snippet (full and compact) gained the same contrast, the memory_message_pop bullet now forbids guessing the inbox's project name, and docs/agent-messaging.md states the distinction.

2.6.3 - 2026-10-09

Fixed

  • Fixed cross-project message tools acting on another project's mailbox. A static MCP client (one that does not forward its lifecycle-hook session id, such as Claude Code's MCP connection) that called memory_message_pop, memory_message_cancel or memory_message_send without workspace and project got whichever project published to the shared active-project slot last, often one open in another harness: the pop claimed that project's message, a cancel without an id cleared its whole outbox, and a send went out under its name as the sender the recipient judges trust by. The three tools now refuse a scope the server inferred, naming the project it would have used and how, and change nothing; an explicit scope or a caller bound to its own hook session proceeds. memory_message_list always reports resolved_scope, scope_source and a hint when it inferred the project, and the SessionStart inbox notice and memory_briefing (new scope field) name the inbox to pass. (#1197)
  • Fixed a purged session coming back. purge-session leaves a tombstone so a late event cannot recreate the session, but live hook ingest creates its session row on a path that never checked it, so the next event for a purged session brought back the session and a new observation. Hook ingest now refuses it, and the delivery is acknowledged as dropped_invalid instead of failing and being retried from the client's spool. (#1194)
  • Fixed purge-session leaving the user's own sentences behind in the cross-project profile's evidence: the profile harvests preference-shaped sentences from prompts into candidate rows keyed to the session, and the purge removed the session but not those rows, so a later pass could still converge a purged sentence into a profile page. The purge now deletes the session's prompt candidates in the same transaction. (#1195)
  • Fixed move-project leaving rows on the old workspace. Pending cross-project messages (addressed to the project and sent by it), entities, page feedback, session purge tombstones and the profile evidence were not re-stamped with the project's other tables, so the moved project lost its pending mail, its entity retrieval and open feedback findings, and a purged session could be recreated. They now move in the same transaction. (#1196)
  • Fixed a Codex, Pi, OMP or Grok session going missing once its store held more than 2,000 transcripts. Looking a session up by id read the first 2,000 files in directory-walk order, so the rest came back as not found: ai-memory run dropped the linked session and started a fresh one, and a transcript import failed. The lookup now reads the transcript that names the session id first, then the newest. (#1189)
  • Fixed Crush sub-agent sessions being listed as sessions of their own: backfill, doctor and the automatic session pick of ai-memory run read the Crush sessions table without skipping rows that have a parent, so a sub-agent session could be imported as a separate session or chosen for the resume. The listing now skips them, as session discovery already did. (#1190)
  • Fixed rules titled in a script without ASCII letters (Cyrillic, CJK, Arabic and others) overwriting each other: multi-page consolidation named every such rule page _rules/rule.md, so only the last survived in the live rules list. Each rule now gets a _rules/rule-<hash>.md name derived from its title. Rule titles in Latin scripts are named as before. A restated non-Latin title (case, spacing or punctuation aside) updates its own page. An existing _rules/rule.md is left in place; delete it once its rule is restated. (#1191)
  • Fixed auto-improve accepting a proposal whose confidence is outside 0 to 1: a model that answered in percent (85) cleared the confidence floor and was staged, shown to the reviewer as 8500% and ranked first when sorting by confidence. Such a value is now rejected as confidence_out_of_range and recorded with the other rejected candidates. (#1192)
  • Fixed the curl hook installer (ai-memory-install-hooks) installing event scripts that could not run. Every script sources the shared _lib.sh helper from its own directory or the one above it, but the installer fetched only the per-agent scripts, so the helper was never written and each script stopped at the line that sources it. The installer now also extracts _lib.sh from the verified archive into the install root, beside the agent directories, and refuses to continue when the archive does not contain it. (#1184)
  • Fixed the curl hook installer (ai-memory-install-hooks) disagreeing with the hook bundles about which scripts an agent has. --agent antigravity-cli always failed: it asked for the seven default event scripts, the Antigravity bundle ships four, and the install stopped at the first missing one. --agent claude-code and --agent grok never installed subagent-start.sh and subagent-stop.sh, which their hook config points at, so install-hooks warned that both were missing. The installer now installs the scripts each bundle ships, and a test checks every script-based agent against the repository bundles. (#1185)
  • Fixed Kiro CLI tool calls being stored with no title or content: Kiro's PreToolUse and PostToolUse hooks carry tool_name and tool_input (plus tool_response after the call), and the capture policy already read them, but Kiro was missing from the tool-capture mapping, so every Kiro tool observation reached the store with a generic title and an empty body. They now get the same tool-family title and output summary as Claude Code's. (#1187)
  • Fixed the generated OpenCode, OpenCode 2, OMP, Pi and OpenClaw integrations losing hook events when the server is saturated. They treated every 4xx as a permanent rejection, so a 429 (or a 408 or 425) was dropped instead of spooled, and their spool drain deleted each queued entry it met on such a response. Both now keep 408, 425, 429 and every 5xx for a later drain, as the shell, PowerShell and native hooks already do. Re-run install-hooks --apply for the agent to regenerate its plugin. (#1188)
  • Fixed Claude Code's Windows PowerShell tool being classified as an unknown tool, so its output never reached the session record and [capture] ignore_paths never checked its command. It is now a shell tool like Bash, in the native hook, the server, and the generated TypeScript integrations, and the shell argument reader also follows PowerShell's -Param:value binding (Get-Content -Path:docs\adr\x.md). (#1186)
  • Fixed Zero's sessionEnd never delivering spooled events even with network allowed in Zero's sandbox. Zero kills a hook's detached children when the hook exits, which took the background drainer down with it. For Zero, session-end now drains in-process within a fixed 10-second budget (hard stop at 20 seconds, inside Zero's 30-second hook timeout); events that cannot be delivered stay spooled. Other agents keep the detached drainer. (#1172)
  • Fixed generated TypeScript integrations (OpenCode, OpenCode 2, OMP, Pi, OpenClaw) flashing a console window on Windows. discoverRemoteIdentity spawned git config --get remote.<name>.url without windowsHide. The spawn now passes windowsHide: true, as the repo-root probe's two git spawns already do (#1169).
  • Fixed AI_MEMORY_LLM_REASONING_EFFORT being ignored for claude-haiku-5-5: the Anthropic provider dropped effort and thinking fields for every Haiku model because Haiku 4.5 rejects them, so Haiku 5.5 always ran adaptive thinking at its default medium effort. Haiku 5.5 and later now receive output_config.effort, adaptive thinking, and thinking: disabled for none, like the other current Claude models (#1180).
  • Documented that Anthropic currently refuses Claude Free/Pro/Max subscription tokens from third-party apps, so anthropic-oauth usually fails with a 429 rate_limit_error on every request. That is a policy refusal, not a real rate limit; the docs now point to the anthropic provider with a Console API key or another provider. Also corrected the claim that claude setup-token writes CLAUDE_CODE_OAUTH_TOKEN: it only prints the token, which you export yourself. (#1170)
  • Documented that Zero runs hook commands inside its sandbox, whose default policy blocks network (loopback included) and writes outside the workspace, so hook events stay in the spool instead of reaching the server. The install-hooks --agent zero and setup-agent notes and the Zero install docs now name the global ~/.config/zero/config.json sandbox keys that allow delivery, their tradeoff, and the ai-memory hook-drain fallback. (#1172)
  • Fixed a global ~/.codex/AGENTS.md reaching the cross-project profile as the user's own words. Codex injects its instruction files as a user message headed # AGENTS.md instructions for <cwd> when a project AGENTS.md contributes, and # AGENTS.md instructions with no directory when only the global file does. The transcript import recognized only the first heading, so backfill stored the global file as a user prompt in every project, and the profile harvest admitted its "always"/"never" lines, which repeat across projects by construction. The import now recognizes both headings. (#1173)
  • Fixed the profile harvest reading text the user did not write as the user's words: a user-prompt observation that is a harness turn, such as Claude Code's <task-notification> blocks or Codex's injected instructions, is now skipped, and ai-memory's own routing and recalled-history blocks inside a prompt are left out, including a block the 16 KiB prompt cap cut before its end marker. profile rebuild also skips Codex instructions an earlier version stored as prompts, but it keeps the candidates already harvested from them, so those lines can still show as waiting in profile review. An entry an earlier version already admitted, or one those candidates reach later, stays until ai-memory profile forget removes it. (#1173)
  • Fixed Cursor tool calls being stored with no title or content: Cursor's own hooks, and the Claude Code hooks Cursor also runs, send Claude's tool_name/tool_input fields, but Cursor was missing from the tool-capture mapping, so every Cursor tool observation reached the store empty. They now get the same tool-family title and output summary as Claude Code's.
  • Fixed every Cursor window adding a session to memory before any conversation: Cursor fires sessionStart for its empty draft composer with the placeholder id empty-state-draft and no workspace, and each launch appended another observation to that one empty session. The server now acknowledges and drops that placeholder.
  • Fixed profile review listing page metadata such as Status:** Accepted as habits waiting for promotion: a curated page's statement was its first line of prose, which on an ADR-style page is a **Status:** Accepted or **Date:** field. Statements now skip fields with an ADR metadata label (Status, Date, Deciders, Supersedes and the like) whatever their value; a labelled rule such as **Package manager:** pnpm is still a statement. Fields harvested before the fix no longer reach profile review; an entry already admitted to the profile stays until ai-memory profile forget. (#1175)
  • Fixed install-mcp --client claude-desktop refusing to run on Linux: Anthropic ships Claude Desktop for Linux as a beta, and the command now writes $XDG_CONFIG_HOME/Claude/claude_desktop_config.json (default ~/.config/Claude/) instead of requiring --config-file.
  • Corrected the Claude Desktop documentation: Linux is an Anthropic beta for Debian-based distributions with its config at ~/.config/Claude/claude_desktop_config.json (pass it to install-mcp --client claude-desktop with --config-file), the Code tab's local sessions are already captured by the Claude Code hooks, and Cowork capture is not claimed. The desktop-app research notes record what was verified live and what is still open. (#878)
  • Fixed ai-memory-wikisync committing its per-clone export state: the git add of the destination that export --apply prints also staged .ai-memory-wikisync/state.json, so two clones would conflict on it and trust each other's baselines. The state directory now ignores itself with a .gitignore of *. A state committed by an earlier export needs one git rm -r --cached <dest>/.ai-memory-wikisync. (#1162)
  • Fixed bootstrap reading a rules file twice when it is reachable under two names: on a case-insensitive filesystem claude.md opened CLAUDE.md again, and on any system a CLAUDE.md linked to AGENTS.md did the same. The doubled text inflated the token estimate, could push every source over the budget (no input sources selected), and repeated the rules in the prompt. Only names the directory lists are read now, and identical text is read once. (#1168)
  • Fixed backfill and managed runs not finding a Claude Code transcript when the session's directory contains anything but letters, digits and /: Claude Code names a project's folder by turning every other character into -, but ai-memory replaced only /, so every Windows path and any path with ., _ or a space missed it, and the bounded fallback scan could run out on a large store first. The folder name now follows Claude Code's rule, every project folder is probed for the exact transcript before the scan, and a subagent's subagents/*.jsonl sidechain, which carries its parent's session id, is never taken for the session's transcript. (#1167)
  • Fixed the macOS menu bar companion's Settings… item opening the Settings window behind the frontmost app, so clicking it appeared to do nothing. The item now activates the app before opening the window, as Show Status… already did. (#1161)
  • Fixed ai-memory run creating an empty twin project when a repository's folder name differs from its git remote (a new-space-game checkout of github.com/acme/unknown-system) and its project predates recorded identities: the run asked for the remote-derived name alone, so it missed the folder-named project, and once hook capture claimed that project every later run failed with project '…' is ambiguous. The launcher now sends the checkout's remote identity with the derived name and the server routes the run by identity, as capture does — which also lets runs in an already-affected repository open again. A user whose run cwd sat inside a project they could not write got a silently created twin project; they now get a 403. (#1182)

2.6.2 - 2026-10-08

Fixed

  • Fixed automatic handoffs carrying harness scaffolding as the user's request: a user turn that opens with a markup block whose opening tag has no attributes (for example the <task-notification> block Claude Code delivers when a background task, agent or workflow completes) no longer becomes the handoff's Started: / Last: summary, its Continue from: line, or its Unresolved question: line. The handoff now uses the last prompt the user actually wrote, and falls back to the no-prompt summary when every prompt in the session is such a block. Prompts that only contain markup later in the text, pastes with attributes such as <pasted_content id="…">, and stored observations are unchanged. (#1158)
  • Corrected the profile rebuild guidance in docs/cross-project-profile.md and the command's help: it never removes an entry, and a candidate recorded by an earlier version keeps the classification it was recorded with, so entries 2.6.0 admitted under the looser rules survive the upgrade and a rebuild. The doc now gives the upgrade path, profile review then profile forget on each entry to drop. (#1155)
  • Fixed a workspace/project label passed as the project failing to resolve (for example workspace: "default", project: "myorg/myproject"): when no project has that exact name, the label's own workspace and project are used, under the same access checks as passing them separately. A write never creates a project whose name contains / any more; it is refused with the hint instead. (#1152, #1154)
  • Fixed every Pi turn failing with a 400 on Moonshot/Kimi models: the generated Pi extension registered the MCP tools with schemas containing $ref, which Moonshot rejects. Its bridge now requests the ?flavor=moonshot schemas, with every reference inlined, which any provider accepts. Re-run ai-memory install-hooks --agent pi --apply to regenerate the extension. (#1157)

2.6.1 - 2026-10-08

Fixed

  • Fixed an explicit-scope miss giving no hint when an agent passed a workspace/project label as the project (for example project: "myorg/myproject" with the default workspace): project names never contain /, so the error now names the separate workspace and project arguments to pass. (#1152)
  • Fixed the cross-project profile rewriting a settled entry whenever new evidence merely agreed with it: a statement now changes only when the ruling or its scope does (the LLM merge reports changed), so the line every project receives stays put. The profile digest also moved to the front of the session-start payload, ahead of the handoff and brief that change every session, so it stays in the harness's cacheable prompt prefix; its footer now asks the agent to open an entry before relying on its one-line summary; and topic grouping no longer compares every candidate with every other, which grew quadratically on long-lived installs. See docs/design-cross-project-profile.md §11. (#1000)
  • Fixed hook observation bodies being capped before the sanitizer ever saw them: excerpt extraction (tool: … bodies, user prompts, notifications, post-compaction summaries, extension bodies) applied its 2 KB / 16 KB ceilings at parse time, so a secret straddling the cutoff was cut in half and the surviving prefix, too short to match a pattern, was stored unredacted — the body-side twin of the #980 title-hint leak. The per-event caps (payload::durable_body_cap) now run at the ingest funnel, applied to the already-scrubbed text, mirroring Sanitized::new and the #1109 feedback-reason fix. The stored body shapes and byte limits are unchanged. (#1114)
  • Fixed the native hook's client-side spool cap truncating oversized lifecycle bodies (user prompts, notifications, post-compaction summaries) before any scrubbing: the capped field is now scrubbed with the built-in sanitizer first and truncated second — the order assistant_capture and the #980 / #1109 fixes established — so a straddling secret can no longer reach the spool (and through it the server) as an unredacted fragment. (#1114)
  • Fixed the managed-workstream ledger capping event content before the sanitizer ever saw it: the 64 KiB per-event cap in hooks::workstream::sanitize_events truncated first, so a secret straddling the cutoff was cut into a prefix too short to match any pattern and persisted unredacted in the immutable raw segment — the same ordering bug as the #980 title-hint and #1109 feedback-reason leaks. The cap now runs after scrubbing, the server-generated checkpoint and losses boundary events get the same ordering, and the store's finish_run scrubs with the caller's sanitizer before its own 16 KiB bound as defense in depth. The stored shapes and byte limits are unchanged. (#1113)
  • Fixed SessionEnd replacing the session page an agent wrote itself through memory_write_page with session_id: the rule-based summary overwrote it on every substantive session end, and the opt-in SessionEnd worker's skip never applied, since it read the page after that overwrite and compared an observation count that the agent's own tool call, the Stop and the SessionEnd always advance. A session page carrying consolidated_by: agent is now kept by both, and by the PreCompact and PostCompaction checkpoints, which rewrote it the same way. The write no longer stamps observation_generation. (#1138)
  • Fixed ai-memory run claude resuming the launch session forever after a /clear: Claude Code continues a cleared conversation in a new session, and the workstream now follows it — the newest one after several clears — so the next launch resumes where the work actually went and that transcript, not the abandoned one, is imported. Only a transcript that records the /clear command, names this launch's session as its origin, and stays in this checkout (or a directory below it) is followed. (#1135)
  • Fixed ai-memory backfill failing on every Antigravity CLI session: it and ai-memory run now import the session's user prompts from agy's history.jsonl, keeping only lines whose conversation and workspace match the checkout. (#1134)
  • Fixed the hook spool charging a spooled event's retry budget while the server was unreachable: an endpoint-level delivery failure (connection refused, timeout, DNS — the existing Unreachable classification) no longer increments attempts, so a total outage no longer deletes the oldest events at roughly one per max_attempts drain passes while the server is down. Post-connect failures (a server that answers with 5xx or a protocol error) keep the previous charging semantics, and the 10,000 file cap and 7-day spool TTL bounds are unchanged. Also corrected the stale hooks/_lib.sh comment that claimed the backlog is drained at session boundaries only (a piggyback drain also runs after any successful 2xx POST). (#1121)
  • Fixed the PowerShell hook bundle silently losing every capture event during a server outage: the .ps1 path now spools an undeliverable POST (connection failure, timeout, or 5xx) to the same <data_dir>/hook-spool/ on-disk contract the shell bundle, the native hooks, and ai-memory hook-drain share — same <ms>-<pid>-<seq>.json entry names, same SpoolEntry JSON — and, like ai_memory_post_hook, kicks a detached bounded drain (≤64 entries) after the next successful delivery, retires entries on a 2xx or a terminal 4xx, and never spools a routed-repository or externally-owned capture event. The PowerShell POST also mints an idempotency ingest_key before its initial attempt and keeps it on the spooled replay, so an ambiguous delivery that committed server-side is discarded on replay instead of double-ingested. (#1122)
  • Continued the memory_explore provider-body redaction (#1103) to memory_query(answer=true): when answer synthesis fails, the answer_unavailable note and the server warning now carry only the redacted class/status summary (for example class=provider status=400) instead of the error's Display, which for a provider failure includes the upstream response body. The default path (no answer) is unchanged. (#1132)
  • Fixed Hermes post_tool_call captures losing the tool result and outcome. Hermes nests the tool result at extra.result and mirrors the call status at extra.status; the extractor previously read only top-level tool_output/tool_response/result, so every Hermes observation body read (no output captured) and the outcome stayed unknown forever. Output now comes from extra.result (falling back to extra.error_message), and extra.status of ok/error maps to success/error. Recognized Hermes tool names that execute code or reach the web (execute_code, browser_exec, browser_navigate, web_extract, web_fetch, delegate_task) classify as non-file instead of unknown, so their output is no longer dropped by the unknown-family body shortcut. Payload shape verified against Hermes agent/shell_hooks.py on 2026-10-07. (#1123)
  • Fixed the Pi extension sending notifications/initialized as a JSON-RPC request with an id: JSON-RPC 2.0 and MCP notifications do not carry an id and expect no response body, which caused the server to emit -32601 method-not-found warnings on every Pi launch. The generated extension now uses mcpNotify to send notifications without an id and accept empty successful HTTP responses without failing JSON parsing. (#1136)
  • Fixed the shell (hooks/_lib.sh) and PowerShell (hooks/lib/ai-memory-hook.ps1) hooks dropping live events and deleting spooled entries on a 408, 425 or 429 response, which they treated as a permanent rejection. They now keep those events queued like other transient failures, matching the native spool (429 retries for free; 408/425 retry under its attempt budget since #1092), and a drain pass stops without deleting queued entries. (#1146)
  • Fixed expired pages still showing in page_links, related_walk and cross_project_edges: the web link panel, memory_read_page include_related, and the web / /api/v1 graph view now hide a page (and any edge touching it) once its TTL has passed. Search, recent, briefing, and graph_neighbors_for_project already applied the retrieval TTL (expires_at is null or still in the future); these three queries only filtered is_latest. An expired neighbour is neither returned nor walked through. Exact-path reads of an expired page are unchanged. (#1141)
  • Fixed the session-aware MCP bridge failing on Docker-wrapper installs: the wrapper sent mcp-bridge to the helper container, where the registered 127.0.0.1 server URL is unreachable, so every session started without ai-memory's MCP tools. The wrapper now runs mcp-bridge through its checksum-verified native host client, which also repairs existing entries. (#1147)
  • Fixed the cross-project profile promoting a single project's "always" / "never" / "by default" sentences as general preferences: only an explicit cross-project scope ("in all my projects", "every project", "everywhere", and the Portuguese equivalents) now waives the min_projects threshold, in both the zero-LLM detector and the LLM classifier prompt. (#1148)
  • Fixed explicit in-session "consolidate this session" requests reaching the server's model: the routing instructions and the managed learning and durable-pages skills now send them to the agent route (memory_read_session_observations, then memory_write_page with session_id) and give the agent the server's multi-page layout (session, concept, decision, gotcha and rule pages, at most five). memory_consolidate stays for other sessions and headless runs. (#1140)
  • Fixed the "project 'X' is ambiguous" refusal giving no clue which projects collided. It now names each project that answers to the name and the key it answers by (its name, its canonical key, or its legacy key), so an operator whose undeclared clone created acme-widget with widget as its legacy key can see which project to purge or rename. A restricted project is reported as "a restricted project", never by name, matching the project list. (#1144)
  • Fixed the cross-project profile learning from agent-written prompt text and counting one fanned-out brief as several projects: sentences that read like agent output (markdown bold, file.ext:line references) are no longer candidates, the same sentence arriving in several projects within an hour counts as one project toward min_projects, and one sentence backs at most one entry. (#1148)

2.6.0 - 2026-10-07

Added

  • Added the ai-memory-wikisync companion (slice 1 of the accepted #986 team-wiki sync plan): a standalone read-only export tool that mirrors explicitly allowlisted page families from a running server's /api/v1 surface into a directory inside a project repository. plan is always a dry-run; export writes only with --apply, stores per-page hashes and ETags in a single local state file, transports the server's canonical page projection verbatim without forging frontmatter, refuses files edited locally since the last export (with a diff summary) unless --force is passed, never deletes anything, and never runs git — it prints the commands instead. Bidirectional sync, deletes, and conflict handling remain tracked in #986. (#986)
  • The openai-compat provider now sends X-Request-Id with every chat attempt, carrying the logical operation id shared by all attempts of one operation — including the strict-to-tolerant fallback. A gateway that records the header (for example as req=<id>) can now correlate every attempt of the same operation with the engine's own logs. A static x-request-id entry in AI_MEMORY_LLM_HEADERS is refused at startup for this provider; the official openai and opencode providers are unchanged (opencode keeps its x-opencode-session contract). (#1104)
  • Added --require-server to ai-memory run (plus run.require_server in config.toml and AI_MEMORY_RUN_REQUIRE_SERVER=true) to restore fail-closed launches when the server is unreachable: the reachability probe still runs, and an unreachable server returns the existing "could not reach …" diagnosis — including "the agent was not started" — without starting the harness. (#1112)
  • Added bounded install-wide identity and component-safe path routes in the exact operator-home .ai-memory.toml. Exact hostful repository identity wins the longest lexical path route, local non-home markers stay authoritative, aliases reuse the existing identity-bound marker wire, and malformed or ambiguous route maps fail closed across native, shell, PowerShell, generated TypeScript, and OpenClaw clients. Recognizable malformed route syntax is rejected before fallback, while malformed root-only settings retain legacy parsing. Generated clients preserve marker versus repo-root provenance, and OpenClaw keeps local git identity discovery alias-gated. The default identity style remains unchanged. (#1033)
  • Added bounded local .ai-memory.toml aliases = ["former-name"] routing for existing project-coordinate compatibility. Native, POSIX shell, PowerShell, and generated TypeScript hook clients forward the same validated aliases for capture and SessionStart handoff lookup; resolution remains workspace-local, requires the checkout's matching hostful git-remote identity, preserves the existing UUID, and permits only write-authorized canonical-name promotion. Aliases are not persisted, and the default naming style is unchanged. (#1033)
  • Added read-only project-coordinate diagnostics to ai-memory doctor and the root-gated GET /admin/project-coordinate endpoint. The report classifies exact, canonical-compatibility, legacy-compatibility, missing, and ambiguous names; shows safe local marker/repository evidence and rename eligibility; and reports collisions without creating, claiming, renaming, or exposing remote credentials or marker paths. Ambiguous results keep local session evidence and mark server capture counts unavailable; any returned preferred project is contextual only, never a unique resolution. (#1033)
  • Added the cross-project profile (docs/cross-project-profile.md, design in docs/design-cross-project-profile.md): a small set of profile/ pages recording how the user usually works, delivered to every project and every harness at session start as a fenced, byte-budgeted digest of defaults that rank below the user's instructions, the repository's rules file and the project's own memory. It is on by default for a single-operator server, shared across every workspace, and off on a multi-user server until [profile] enabled = true, which gives each operator a private, restricted profile; share = "global" or "workspace" makes a team profile that is read-open and write-gated like _global, admits a habit only on evidence from at least two operators, and is presented as team defaults. The digest filters entries by the stack the project's activity shows, stays byte-identical between sessions, and gives a project with no memory yet a larger baseline. memory_write_page and memory_delete_page accept scope: "profile", memory_query returns profile entries in global_scope_hits, a marker's [profile] contribute / consume keys opt a project out of contributing or receiving (an opt-out stays recorded until a marker says otherwise), and ai-memory profile status | list | show | forget | review | rebuild inspects and curates it, while ai-memory profile apply writes the entries that fit a repository into a managed block of its rules file (--remove takes it out), never touching anything outside the block. Entries are learned on their own: a server pass, run after each session ends and hourly, harvests the user's own preference-shaped prompts (never tool output), the project's curated pages and its stack signals, and converges them across projects into entries once the user states a choice as general or it appears in min_projects projects, the newest ruling winning. With an LLM provider the pass classifies and restates entries through JSON-schema calls and falls back to the zero-LLM path on any error. It never rewrites an entry edited by hand, and an entry removed with profile forget stays removed until the user states it again. (#1000)
  • Added ai-memory resume --search with immediate type-to-filter name search, and resume --all, which keeps the previous cross-checkout picker: every linked checkout, the current one first, each listed with its own fully paginated query. (#1039)
  • Added install-hooks --agent claude-code --scope project, which writes the hook configuration to the checkout's gitignored .claude/settings.local.json (where Claude Code reads it: the git root, or the launch directory on Windows and when the repository root is the home directory) instead of the user-level settings.json, so capture is opted in per repository. The installer warns when that file is not ignored by git, refuses the inline-token fallback for it, notes when the other scope also carries ai-memory hooks, and keeps its backups under the data dir instead of the checkout; uninstall run inside the checkout sweeps it and reminds you to do the same in other project-scoped checkouts; upgrade and the ai-memory run auto-wire no longer add user-level Claude Code hooks on top of a project-scoped install. (#1034)
  • Added explicit MCP behavior annotations (title, read-only, destructive, idempotent, and open-world hints) to all 23 tools. Reordered the MCP server instructions so scope selection, untrusted-memory handling, deliberate cross-project broadening, capture policy, and handoff safety remain complete within the first 2,048 characters used by truncating clients. The longer per-tool routing reference remains available to clients that accept it. (#920)
  • Added env-only named launch profiles for ai-memory run --profile NAME via [run.profiles.<name>.env] in config.toml. Profile values feed the same native-session, auto-wire, and child-process environment path as --env; precedence is profile, then --env-file, then repeated --env. Unknown or invalid profiles fail before a workstream lease or child process; an unknown name lists the defined profiles or prints a table to paste. (#922; error message, tests and docs from #1031 by @geeksilva97)
  • Added cross-project attribution for native lifecycle file-tool events whose harness keeps reporting the parent session's cwd. A bounded, recognized absolute target reroutes only when every path proves the same repository or marker boundary; destination capture exclusions and server profiles are resolved before the event is spooled. Relative, mixed-project, unsupported, and arbitrary non-project paths keep the original route. (#932)
  • Added serve --enable-api / AI_MEMORY_ENABLE_API=true to mount the authenticated read-only /api/v1 surface without the browser UI. Existing --enable-web behavior is unchanged and still includes the API. Documented the acceptance and safety boundary for a repository-backed team-wiki sync companion. (#986)
  • Added [handoff].claim_on_session_start (default true - unchanged behavior). When set to false, SessionStart no longer claims a pending handoff automatically; it renders a non-consuming notice naming the exact handoff_id, from_agent, and age instead (mirroring the existing inbox notice: metadata only, never the stored summary/open-questions/next-steps text), and the agent or operator picks it up explicitly with memory_handoff_accept. Fixes an unrelated next session, a different harness, or a non-interactive launch silently consuming a baton meant for a specific session. Server-wide for now; a per-project override is left for a follow-up change. (#1030)
  • Added a default-quit prompt after a successful interactive managed run so the operator can re-run the current harness or switch to another installed harness in the same workstream. Utility, failed, interrupted, non-interactive, and custom---executable runs keep their existing exit behavior. Resolves #909. (#975)
  • Added ai-memory backup-agents and dry-run-by-default ai-memory restore-agents commands for host agent configurations, skills, plugins, and project instructions. Archives are written atomically with private Unix permissions, MCP config text is sanitized by default, and restore bounds archive input, verifies manifest checksums and completeness, refuses duplicate or symlink-escaping targets, and keeps overwritten files as timestamped backups. Other asset types are copied verbatim and may contain secrets or active instructions/code. (#962)
  • Added a NixOS module (nix/nixos-module.nix, exposed as nixosModules.default) with a dedicated non-login system user and a hardened NixOS systemd sandbox. A minimal typed set plus freeformType renders declarative config.toml; values land in a world-readable store path, so secrets (including llm_headers) use ageSecret, sopsSecret, or environmentFile. Non-loopback binds require a secret source. The module includes opt-in web, firewall, and resource settings plus eval and privileged container checks. Flake packages are exported for x86_64 Linux, aarch64 Linux, and Apple Silicon macOS; Intel macOS remains available through the release tarball and Homebrew rather than the pinned Nixpkgs revision. (#989)
  • Added ai-memory doctor reporting for Claude Code's default native memory/ store for the current repository: location, file count, and whether the nearest marker's ignore_paths would exclude a read. Repository-root resolution keeps worktrees and subdirectories on the same report, and the output warns that shell/PowerShell compatibility hooks do not enforce the exclusion. Resolves #1003. (#1005)
  • Added per-event outcomes to /hook/batch acknowledgements and process-lifetime ingest counters for stored events, replays, recovery, ignored endings, collisions and failures. Legacy acknowledgement fields were preserved. (#1015)
  • Added receipt outcomes and delivery counts to the companion relay. Opening a schema 1 queue now upgrades it atomically to schema 2; older relay binaries refuse schema 2, so downgrade requires a pre-upgrade backup with all queue users stopped. Pending order and existing delivery metadata were preserved. Receipts now retain their first known outcome. (#1015)
  • Added ai-memory list-projects [--workspace] [--json]: a plain, scriptable listing of every workspace/project pair the server knows about, sorted by workspace then project. Previously the only way to see this from outside an interactive show session was to call GET /api/v1/projects directly. (#1022)
  • Added scope and policy preflight to ai-memory hook --check-capture, including event eligibility, sanitized routing hints and server-remapping information. Partial scope, excluded events, rejected profiles and oversized hints refuse capture. Inspection neither ingests events nor claims handoffs. (#1019)
  • Added the latest scoped consolidation job's state and attempt count to HTTP session summaries and MCP memory_read_session_observations. The field is null when no job exists; provider errors remain private and existing session ownership filters are preserved. (#1018)
  • Added incremental paging to scoped HTTP recent reads with updated_since and an opaque, scope-bound cursor. Queries preserve equal-time ordering, exclude expired and superseded pages and recheck authorization on each request. Calls without incremental arguments retain their existing array response. Added a generic MCP/HTTP integration guide. (#1017)
  • Added machine-authenticated GET /identity, available with the web UI disabled, and ai-memory doctor diagnostics for caller identity, external capture ownership and sessions with multiple capture sources (native events or distinct extensions, including backfill). Mixed-source counts preserve scope, session ownership and time-window filters; older servers leave unsupported fields unknown. (#1016)
  • Added shared capture privacy checks to the relay and the conversation importer, run before an event is hashed, persisted, previewed or delivered: they refuse credential-shaped native identities, scrub nested credentials while keeping numeric token counts, and drop locally a relay item queued before them that fails them, so it no longer holds its session. (#1072)
  • Added [handoff].create_on_session_end (default true - unchanged behavior). When set to false, unmanaged session ends write the session summary page and enqueue consolidation as usual, but skip creating an automatic open handoff for the next session; OpenCode turn checkpoints refresh the page without a baton too. Explicit handoffs created via memory_handoff_begin and managed workstream runs remain unaffected. Server-wide: applies to every operator on this server. (#1043)
  • Added bounded, sanitized native-record and adapter correlation provenance to managed-workstream search/tail results and the CLI's text/JSON output, with compatible optional fields and sanitization of legacy provenance on reads. (#1057)
  • Added Portuguese history phrases to the existing opt-in session-recall router ([retrieval] query_intent), including accented and unaccented spellings of "última sessão", "onde paramos ontem", and "decisão anterior". Word boundaries and explicit history phrases keep technical session queries unchanged; routing remains off by default. (#1056)
  • Added shared, typed and bounded entities, abstract, and relations metadata to MCP memory_write_page and admin /admin/write-page, plus bounded MCP kind while preserving the admin's legacy kind adapter. Kept existing request shapes, entity normalization, authenticated attribution and the wiki write pipeline. Documented whole-page replacement, raw-input metadata bounds and trimmed relation scope components. (#1055)
  • Added structured results for the existing auto-improvement eval execution in review reports, staged run metadata, and accepted proposal sidecars, with server-generated eval IDs, evaluated request/body digests, checker invocation identity, observed success/rejection/failure/timeout, and sanitized bounded reasons. Report deserialization discarded supplied eval observations. (#1053)
  • Added grizzybot as a recognised agent (wire value grizzybot, alias grizzy-bot), so GrizzyBot's lifecycle events are attributed to it instead of other and its tool calls are captured as closed-schema tool families, like other Claude Code-shaped agents. GrizzyBot posts to /hook and /mcp itself, so there is nothing to install. Adds migration V72 to extend the sessions.agent_kind CHECK constraint. (#1036)
  • Added GitHub Copilot CLI as a hooked agent: install-hooks --agent copilot-cli writes $COPILOT_HOME/hooks/ai-memory.json (default ~/.copilot/hooks/ai-memory.json) with Copilot's flat, matcher-less entries and PascalCase event names, so Copilot sends its VS Code/Claude-compatible payload. Ten events — Claude Code's nine plus PostToolUseFailure — with native commands enforcing capture exclusions; tool output is read from tool_result.text_result_for_llm and the outcome from tool_result.result_type. The SessionStart hook delivers the prior session's handoff through Copilot's top-level additionalContext (ported from #1069). There is deliberately no --scope project. The bare install-mcp --client copilot alias keeps meaning VS Code Copilot; ai-memory run copilot follows separately. (#1040)
  • Added install-mcp --client copilot-cli, which merges ai-memory's remote entry (type: "http", url, bearer headers, tools: ["*"]) into the root mcpServers map of $COPILOT_HOME/mcp-config.json (default ~/.copilot/mcp-config.json), preserving other servers. install-hooks --agent copilot-cli without --server-url now reads the server URL and bearer back from that entry, and uninstall removes it. (#1040)
  • Added an optional session_id to memory_write_page, so an agent can write a session's page with its own model and keep it traceable to the session. The session must belong to the project the page is written to, and the page records it as session evidence. A page that cites a session does not overwrite a pinned page. Writing sessions/<id>.md also stamps the frontmatter memory_consolidate writes there (session_id, agent, consolidated, plus consolidated_by: agent), defaults the tier to episodic, applies the same duplicate-title suffix, and marks the session's queued SessionEnd consolidation job completed. (#1038)
  • Added identity_style = "path" | "host_path" marker control for remote-backed project naming. path names from the full repository path without the host (github.com/acme/api → acme-api); host_path preserves the former naming behavior as an explicit opt-out. Captures always route by the full hostful identity, and cross-forge path collisions never merge. (#1033)
  • Added [hook_spool] max_attempts and AI_MEMORY_HOOK_SPOOL_MAX_ATTEMPTS to configure failed drain passes before dropping an event; zero disables attempt-based drops while age and spool size limits remain. (#1092)

Changed

  • Redacted provider response bodies from the error surfaces this change touches: the memory_consolidate and memory_auto_improve MCP errors, the memory_explore degradation reason and warning, the reranker degradation warning, the consolidator's transient-retry warning, the SessionEnd consolidation worker's queue last_error and failure logs, the auto-improve scheduler's claim last_error and tick warnings, and the hook router's checkpoint-fallback warning now carry a stable class/status summary (for example class=provider status=400) instead of the error's Display, which for a provider failure includes the upstream response body. Other error surfaces are unchanged. (#1103)
  • Changed ai-memory run to no longer fail closed when the server is unreachable: a bounded ~2s reachability probe (any HTTP answer counts as reachable, so pre-/healthz servers pass) downgrades the launch with one loud warning instead of aborting it. The harness starts with no server work at all (no lease, link, context, heartbeat, status, finish, or import), lifecycle hooks keep spooling locally with the bounded retention surfaced in the warning, auto-wire installs hooks but defers the MCP registration to the next online launch, sessions resume only through an explicit native selector (no lease means no mutual exclusion), the child runs without AI_MEMORY_RUN_ID/AI_MEMORY_WORKSTREAM_ID so its session cannot be misattributed, and the child's exit code is returned with a local spool-health account. A server that dies mid-run, after the child exits, no longer discards the exit code either: the unimported transcript is downgraded to a warning naming the exact ai-memory finalize-session repair command, while protocol-level failures still fail hard. (#1112)
  • Changed static CLI/MCP project-name resolution to accept an existing repository's canonical hostless path name and its v2 basename compatibility key. Reads remain find-only; an authorized write through the canonical key promotes only projects.name in place, preserving the project UUID and all dependent rows. Project scope manifests preserve the optional full hostful identity/source and derived keys so clean reindex reconstructs the same UUID and lookup behavior; older manifests remain valid. MCP, admin, and managed-workstream writes refresh the manifest after promotion and disclose a post-commit refresh failure for startup repair on both success and any later operation error, retaining the first warning across busy retries and lease wait-out; managed-run caller validation happens before promotion. Ambiguous cross-forge keys and target-name conflicts fail closed. Current clients now explicitly send path for a valid remote and send explicit host_path for the opt-out; a new server deliberately interprets an omitted field as legacy host_path so old clients cannot be silently renamed. Existing compatible projects are promoted only by authorized capture/write and keep their UUID; read-only handoff/static lookups never rename. (#1033)
  • Changed the documented precedence between rules and memory: the user's instructions, then the repository's rules file, then the project's memory, then the cross-project profile. Where no hard rule exists, agents now use what memory already knows instead of asking again; the routing snippet, MCP instructions and durable-pages skill send standing cross-project preferences to the profile (scope: "profile"). (#1000)
  • Scoped ai-memory resume to the current checkout by default instead of every linked project (--all restores the old view) and removed its default result cutoff with bounded server pagination, preserving Left/Right harness selection. Explicit limits now apply after the initial search; against an older server that ignores paging, the picker warns and shows the first page. continue retained its cross-directory behavior. (#1039)
  • Documented FutureInfra as an endpoint for the existing openai-compat provider. (#1026)
  • .github/workflows/nix.yml builds the flake on x86_64-linux for path- filtered PRs and pushes (package smoke + NixOS module eval / sandbox-parity). The aarch64-darwin package smoke and the privileged NixOS container smoke run only on schedule, workflow_dispatch, or a PR labelled nix / full-ci. (#989)
  • Restricted managed-workstream provenance to scrubbed source labels and an explicit scalar metadata allowlist, with 512-byte string bounds. Client event ids beginning with managed-run: are reserved for server checkpoints and extraction-loss annotations. Active-run retries reject changes to an existing event's agent, native session or kind, and retain its first indexed content and provenance across sanitizer changes. (#1057)
  • The SessionEnd consolidation worker leaves a session page alone when the agent wrote it through memory_write_page and no observation arrived after it: the page records the session's observation count as observation_generation, and the worker completes a job whose generation that count reaches instead of replacing the page with the server's provider. A page written before the session's last observation is consolidated as before. (#1038)
  • Required managed-run finish imports and finished retries, and the cancel, heartbeat, native-session link and context-acceptance routes, to pass an authenticated owner check and current project access checks (repeated in finish's writer transaction), so another operator cannot release or rebind a run. Database users required current Write access; trusted proxy operators retained their existing project policy. Active API keys remained valid after human login was disabled. Shared, unattributed runs and late finishes of still-active runs retained their existing behavior. (#1075)
  • Required new managed native-session bindings and imports to preserve the exact original identity within 512 UTF-8 bytes and pass privacy validation before use or persistence. Invalid historical identities were projected as unknown (native_session_id: "") in workstream search without rewriting stored history; the CLI also protected reads from older servers. (#1079)
  • Ordered the default single-page and batch consolidation prompts for prefix-cache reuse: in the user content, the fixed batch text (header + field schema) and the project instructions now lead, followed by the current body / slot and title state, with the session id and the observation dump last. In the single-page prompt the current body is the session's own page body, so the prefix shared across sessions of one project is the system prompt plus the instructions block when present. The system prompt, budget math, clipping, and observation selection are unchanged — blocks were only permuted — and the input budget remains an approximate character estimate (no tokenizer ceiling, no cache hit guaranteed). (#1102)
  • The consolidation and auto-improve operation id is now a fresh UUID v7 generated by the caller of the LLM operation — once per consolidation invocation and once per auto-improve review — and is never derived from the agent's session id, so two operations of one session (or a crash-resumed or re-claimed run of the durable consolidation queue) are distinct operations to the gateway. Crash re-entry is exactly one new operation: a process restart or a queue re-claim mints a new id for the calls it makes. (#1104)

Fixed

  • The CLI's "could not reach …" reachability diagnosis now fires for connect timeouts and generic connect failures, not only an explicit "Connection refused": on Windows a firewall-dropped loopback connect is reported as "operation timed out", which previously fell through to the generic "HTTP request to … failed" message and hid the resolved server URL and AI_MEMORY_SERVER_URL guidance. The hint now says "nothing answered (refused or timed out)". (#1112)
  • Fixed memory_feedback storing an unmatched secret prefix when a reason straddled the 500-character cap: the reason is now scrubbed before the cap is applied, matching the #980 title-hint order. (#1109)
  • Fixed wiki link extraction treating file: and vbscript: as project qualifiers, so [[file:notes/x.md]] indexed a cross-project edge to project file. Both prefixes now match the /web scheme denylist and are not wiki links. (#1110)
  • Fixed hook title_hint extraction leaving a trailing CR on Windows CRLF payloads and keeping embedded newlines on SessionStart, Notification, and PostCompaction titles. First-line splitting now uses str::lines. (#1111)
  • Fixed wiki confinement walks failing on Windows with PermissionDenied (sharing violation) or ERROR_DELETE_PENDING when a concurrent commit or the atomic writer held or rename-replaced a file mid-walk; per-entry walk probes now skip those as transient churn exactly like the NotFound vanish race, while every entry that does stat is still refused when link-like and root-level probes keep their strict semantics. (#1107)
  • Fixed strip_leading_h1 failing to strip setext H1 headings on CRLF line endings. The underline check required every character of the second line to be =, which failed when the line contained a trailing carriage return, trailing whitespace or tabs, or when the document ended at the underline without a trailing newline. (#1108)
  • Fixed generic OpenCode commands selecting integration contracts from command spelling instead of the executable's major version. run opencode now probes the exact executable once and carries the resolved V1/V2 dialect through session discovery, launch, auto-wire, and transcript import; hook and MCP installers use the same strict resolver, while container-oriented setup-agent requires an explicit V1/V2 artifact dialect. Probe and launch share one captured child environment with --env overlays, including Windows key casing. Automatic launch skips an unresolvable OpenCode candidate without hiding other harnesses. Hook/MCP transitions reuse ownership-verified endpoint and token settings across both schemas and reject conflicting generated entries. Migrations remove only entries proven generated by ai-memory, and plugin cleanup never overwrites a concurrent user replacement while restoring an unowned quarantined file. (#1097)
  • Fixed Codex on the Kimi coding endpoint failing every model call with "tools.function.parameters is not a valid moonshot flavored json schema (... detected infinite recursion without termination condition ...)". Moonshot's validator never resolves $ref, so the $defs/$ref pairs Codex forwards from MCP input schemas failed the request at any depth — including the ReasoningTier reference on memory_query and memory_explore and the FeedbackKind reference on memory_feedback — not just at the root. The ?flavor=moonshot dialect now inlines every #/$defs/* reference and drops the emptied $defs table; nested combinators (which Moonshot accepts) and runtime validation are unchanged. ?flavor=bedrock and ?flavor=gemini keep serving $defs/$ref as before. (#1093)
  • Confined wiki reads, writes, indexing, recovery, and lifecycle cleanup to real project-tree directories by refusing symbolic links and filesystem reparse points at namespace roots or descendants. (#1107)
  • Fixed the web page view omitting heading IDs, preventing #anchor fragments and section wikilinks from scrolling to target headings. The Markdown renderer now emits unique, slugified id attributes on heading elements (<h1>–<h6>), disambiguating duplicate headings and stripping punctuation. (#1105)
  • Fixed a later hook event carrying a managed-run id linking its native session to that run without the checks SessionStart applies. It now links only from the run's own project and operator; an event from elsewhere is still captured but no longer repoints the run. (#1082)
  • Fixed completed retries and no-op session endings advancing last_persisted_ms without a durable write. Recovery still advances the timestamp when it commits a new page or terminal effect. (#1015)
  • Fixed interrupted launchers blocking an immediate managed-workstream restart by adding explicit ai-memory run --force-unlock recovery. The server atomically expires and replaces only a lease attributed to the same authenticated operator (or another unattributed lease in single-user operation), while cross-operator eviction remains refused; the command does not kill a native process. (#795)
  • Fixed Unix release archives potentially carrying macOS AppleDouble sidecars outside the native upgrader's strict path allowlist. Release packing now disables sidecar generation, and packaging tests lock the staged top-level layout to the upgrader allowlist so future release changes cannot silently make official archives un-upgradable. (#1029)
  • Fixed ai-memory upgrade rewriting native hook commands with Linux's (deleted) executable-path suffix after replacing its own binary. Every hook renderer now keeps a still-existing literal path, otherwise strips the kernel suffix only when the resulting installed binary exists, and finally falls back to the bare ai-memory command instead of embedding a dead path. The in-process refresh continues using the already-loaded CLI/config scope. (#1027)
  • Fixed native ai-memory upgrade rejecting every current Linux, macOS, and Windows release archive when it encountered the packaging, config-template, or documentation entries shipped beside the binary. The extractor now validates the exact support-file layout emitted by release.yml, ignores those non-runtime files during self-upgrade, and still extracts only the binary and hooks. Archive entry-count and expanded-size caps now complement the existing compressed-download limit. (#1025)
  • Fixed the recommended Linux/macOS Docker-wrapper install-hooks path silently installing shell hooks that cannot enforce client-side capture controls such as [capture] ignore_paths and allowlist mode. The wrapper now uses its existing checksum-verified native host client for hook installation and refresh, while an explicit AI_MEMORY_HOOK_PLATFORM=posix|windows keeps the documented compatibility fallback and warning. Wrapper upgrades refresh that stable native client before rewriting existing hook registrations. (#1002)
  • Fixed managed Codex sessions losing their startup continuity when a shared Codex app-server daemon reports a finished run's stale AI_MEMORY_RUN_ID. SessionStart now recovers only the sole live, undelivered Codex run in the already-authorized repository, checkout cwd, and operator bucket, and links it atomically to the new native session. Active boundary mismatches, concurrent matching runs, cross-project or cross-worktree candidates, cross-operator candidates, and a second linker all fail closed instead of guessing or rebinding a run. (#987)
  • ai-memory doctor now warns when the nearest .ai-memory.toml's [capture] section is invalid. This fails closed today — every file and shell tool event is reduced to metadata until the marker is fixed, nothing leaks — but nothing surfaced that it had happened, so an operator could believe ignore_paths was excluding paths it no longer was. The check is based on the same resolution state the live hook path uses (not just whether the TOML parses), so it also catches a [capture] table that parses fine but is still rejected at compile time, e.g. an unsupported glob character in ignore_paths. (#1021)
  • Fixed the shell and PowerShell hook scripts mishandling a $HOME that ends in a separator: the shell walks missed a marker between the checkout and home, and the PowerShell walks went past home and read a marker above it. Both now match the native ai-memory hook walk. (#1023)
  • Fixed ai-memory run --jail leaving the ssh toggle unchecked when git push goes over SSH but origin's fetch URL does not show it: an HTTPS origin with an SSH pushurl or pushInsteadOf, or a host:path remote using an ~/.ssh/config alias. The interactive offer now also uses ai-jail's own dry-run preflight, so a backend that exists but fails ai-jail's trust checks is treated as unavailable instead of producing a broken offer. (#1024)
  • Concurrent relay queue opens no longer report a locked queue as foreign: journal setup now retries SQLite BUSY/LOCKED within one 10-second budget, rechecking the queue's identity and schema version before each attempt. (#1060)
  • Fixed watcher reindexing racing with writes and batches to the same page by sharing their per-page mutex from disk read through SQLite upsert. Both mutation guards are released before embedding; external editors remain outside this coordination. (#1059)
  • Fixed a managed Claude Code run losing its conversation after /resume on a Claude Code background session. The run used to finish on the foreground session it started with, which held none of the conversation, so the next ai-memory run claude or continue resumed it and Claude opened an empty session. A Claude run now finishes on the background session its own session attached to in this checkout, found from that transcript's sessionKind: "bg" records; another launch's background session is never taken. (#1050)
  • Fixed continue, resume and bare ai-memory run losing sessions launched under a custom native store, such as a second Claude Code config home set with CLAUDE_CONFIG_DIR. Automatic harness selection now scans the store the launch resolves (--env, then the process environment) instead of always the default one. The client records the store of every session it links, at launch or when the run finishes, in client-projects.json. A later launch that cannot find that session and resolves a different store now stops with an error naming both stores and keeps the workstream link, where it used to start fresh and repoint the workstream. --fresh still starts a new session. (#1047)
  • Fixed Claude Code sessions never seeing the "broaden when the current project comes up empty" and "maintained pages are evidence, not authority" rules, because Claude Code truncates MCP server instructions at 2,048 characters. The instructions now open with a self-contained core under that cap, and a regression test keeps it there. (#1035)
  • Fixed the macOS menu-bar app showing only a red status item when the bundled server cannot start (for example a port already in use; #1044 reports OpenCode v2's background service on 127.0.0.1:49374). The menu extra now surfaces a fatal stderr.log line written since the last start, and docs/macos.md documents the port collision and how to move one side. (#1044)
  • Fixed hook-spool drains stalling behind an event that has no session id. /hook/batch reported such an event (anything but a session start) as a failed item, so the drain retried it up to its attempt budget while every event queued behind it waited. The server now acknowledges and drops it, counted as dropped_invalid in the ingest metrics, and keeps processing the rest of the batch. (#1062)
  • Fixed drop_subagent_captures discarding entire top-level Claude Code sessions launched with --agent. An agent_type alone no longer marks a session as a subagent; agent_id and Grok's subagentType still do, so actual subagent filtering remains enabled. Claude Code reports a Task subagent under its parent's session id, so a subagent's events no longer mark that session as a subagent: the parent's Stop, SessionEnd, summary and handoff are kept. (#1041, #1048)
  • Fixed the generated OpenCode 2 plugin ending every tracked session when OpenCode unloads it on idle-location eviction. Unload is not shutdown: the host and its sessions stay alive, so each eviction froze a live session (SessionStart cannot reopen an ended OpenCode session), fabricated a spurious summary page and open handoff, and made the next baton fetch fail with invalid state: an ended session cannot accept a handoff. Plugin unload no longer posts session-end; deletion (session.deleted) remains the end signal and a completed root turn already publishes its continuation checkpoint and baton. Regenerate the plugin with ai-memory install-hooks --agent opencode2 --apply. (#1074)
  • Corrected the [auto_scope] default in doc comments (Config, AutoScopeSettings, the ActiveProjectMode module docs, serve) and the README docs index: they still named single as the default or called per_actor opt-in, but the default has been per_actor since v1.39. (#1065)
  • Fixed gemini-3.8-flash consolidation and lint spending hidden thinking tokens on strict-JSON calls. The Gemini provider now sends thinkingBudget = 0 to 3.8 Flash as it already did for 2.5 and 3.5 Flash (verified live: 0 thought tokens, about half the latency); 3.8 Flash-Lite keeps the field omitted until verified. (#1077)
  • Fixed the secret scrubber missing Gemini authorization keys. AI Studio has issued AQ.Ab… keys instead of AIza… since 2026-05-28, so a bare new key reached capture unredacted; it is now redacted as google_api_key. (#1077)
  • Fixed the OpenCode provider sending every Go/Zen model through Chat Completions. It now selects the published wire API per supported model: Responses, Anthropic Messages, or OpenAI-compatible Chat Completions. Claude Sonnet 5.5 and the legacy Sonnet 4 ID also use Messages. Its default model is now mimo-v2.6-flash, matching the default Go catalogue; Claude models require an explicit Zen base URL. Configured base URLs remain authoritative, with no automatic switching between Go and Zen billing. MiniMax M3/M2.7 and Qwen3.8 Max use Messages on Go and Chat Completions on Zen; Go's Muse Contributor IDs use Responses. Changing the base URL also reselects the transport while preserving configured headers, reasoning, and timeout. (#1080)
  • Fixed [[_global:path]] resolving against the source page's workspace instead of the reserved _global project in the default workspace. A page outside default can now graph-link to standing global pages, and /web points at the same home. Sibling [[project:path]] and explicit [[workspace/project:path]] are unchanged. Legacy rows that stored the project-only form with a NULL workspace resolve when the reserved page is next written. (#1042)
  • Fixed ai-memory run claude / continue exiting with Claude's "That session is running in the background" error when the linked session is a Claude Code background session that is still running and the launch carries any native flag. The launcher now confirms the session is live in this checkout with claude agents --json (only for a transcript recorded in the background) and opens it with claude attach <id>; any failure keeps the native resume. Documented the observed attach-hook behavior, early-exit import boundary and fallback validation limits. (#1052)
  • Fixed resume and workstreams running git status on every linked checkout before listing anything: listing now reads only the checkout's stable fingerprints, so a large working tree or a slow filesystem-monitor hook no longer blocks the picker before its first frame. (#1039)
  • Fixed a store write sent while the writer was shutting down occasionally waiting forever instead of failing with WriterClosed: the writer now closes its queue and drains in-flight commands before it stops.
  • Fixed memory_handoff_list MCP tool calls being counted under writes in client_activity: the read-only inspection tool is now classified as a read in tool_call_is_write. (#1088)
  • Fixed the provider-health wrapper dropping the caller's operation id: its complete_with_operation_id and complete_structured_raw_with_operation_id fell to the trait defaults, so every retry of one logical operation got a fresh id and, for OpenCode, a new session header on every call. The wrapper now forwards both. (#1085)
  • Fixed backfill failing with a non-root key on a multi-user server. Its emptiness check used the root-only /admin/sessions/by-agent route, so every developer's own key got 403; on a 403 it now asks the grant-checked GET /api/v1/workspaces/{ws}/projects/{p}/sessions instead. (#1086)
  • Fixed the sanitizer missing uppercase *_KEY_ID, *_PASSPHRASE, *_SIGNING_KEY, *_PEPPER and *_SALT assignments (e.g. an S3 …_ACCESS_KEY_ID, an auth token pepper) and dotless base64 JSON tokens such as Cloudflare tunnel tokens (eyJ…, 40+ characters); lowercase code identifiers, SHAs, UUIDs and short eyJ fragments are left alone. (#1084)
  • Fixed memory_delete_page bypassing per-user slot namespace routing and access checks when [slots] per_user is active. Non-admin callers deleting a generic slot path (_slots/current-focus.md) are now redirected to their personal slot namespace instead of removing the shared project-wide slot, and deleting directly inside another operator's slot namespace is rejected with an invalid request error. (#1083)
  • Fixed the web page view showing a [[wikilink]] inside a raw HTML block (an HTML comment, a <div> block) as generated link markup such as [notes/foo](w/default/scratch/p/notes/foo.md). HTML blocks render as escaped source text, so the wikilink now stays as the page wrote it. (#1087)
  • Fixed wikilinks dropping #anchor and ?query suffixes. [[page#section]] and [[page#section|label]] lost the fragment when rewritten to standard Markdown links (for example on OKF export), leaving the link pointed at the page root. The rewriters now keep the suffix, percent-encoded where needed so a space or parenthesis in it cannot break the link. In /web the suffix is kept on the href; the page view does not emit heading ids. (#1089)

Security

  • Fixed GHSA-7qj3-7wqw-m5w6: in multi-user mode a database user without a write grant on the reserved global preferences scope (default/_global) could write it, and its pages are unioned into every user's queries, including restricted projects. Writing it now needs root or an explicit write grant on _global, whichever way the scope is named; creating it grants nothing, reads stay open, and installs without database users are unchanged. Reported by @Josehbr.

Docs

  • Corrected the Codex support matrix to describe managed-run recovery from a stale shared-daemon run id; --no-daemon remains an optional diagnostic and isolation switch. (#987)

2.5.2 - 2026-10-01

Added

  • Added ai-memory run --jail[=TOGGLES] and --no-jail, and an interactive checklist after the --yolo ai-jail offer, to choose which ai-jail credentials and capabilities the jailed session gets. Bare --jail re-runs inside ai-jail without asking, using smart defaults: every credential present on the host (~/.config/gh, ~/.aws, ~/.kube, ~/.config/gcloud, ~/.docker/config.json), SSH when origin is an SSH remote, and worktree metadata in a linked worktree; host capabilities (docker, gpu, display, pictures, tailscale) stay off, and anything else is left to the user's own ai-jail config. --jail=github,aws,no-mise is exact: it passes every checklist row it does not name as --no-X, so a global ~/.ai-jail cannot add to it (all and none also work); the checklist likewise passes unchecked rows as --no-X. --jail works without --yolo and in scripts, and fails instead of running unjailed when ai-jail is not usable. Only toggles the installed ai-jail advertises in its --help are offered or passed; the credential mounts need ai-jail 2.5.0. ai-jail's security switches (seccomp, landlock, private-home, …) are never accepted. --no-jail skips the offer while keeping the --yolo warning. A project .ai-jail in the launch directory replaces the checklist and the bare---jail defaults (ai-jail loads it under its own trust rules; --jail=… still applies on top), and every jailed re-run now passes --no-save-config, so ai-jail no longer writes ai-memory's --network / --agent-state / credential flags into the repository's .ai-jail.

Fixed

  • Fixed the shell and PowerShell session-start hooks for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Command Code, Devin, and Antigravity CLI not sending the marker's [briefing] keys (briefing, briefing_budget) on the handoff request, so a repository with inject_on_session_start = true got the handoff without its compiled brief on script installs (Docker wrapper, setup-agent). Only the native ai-memory hook command and the Kiro CLI and Kimi Code scripts sent them; all now match docs/marker-file.md. (#998)
  • Fixed a wiki checkpoint leaving .git/index behind the commit it made. A path-scoped checkpoint wrote the index file only once every 50 commits, so between writes HEAD and the working tree held the new page while the index still named the old blob, and git status from outside the server showed every checkpointed page as MM; a checkpoint with nothing to commit left a stale index the same way. The history itself was always correct. Not specific to Windows. (#983, #1006)
  • Fixed a wiki checkpoint failing instead of retrying when a full walk listed a file that was gone by the time it was read (libgit2's Os-class "failed to read file into stream", e.g. an atomic writer's temp file renamed away mid-walk). It now takes the same bounded racy-read retry as a file changed mid-write; unrelated I/O errors still fail fast. This was also the source of an intermittent concurrent_commits_queue_instead_of_failing CI failure.

Security

  • Fixed GHSA-gf78-hf8g-vffm: memory_read_page with include_related returning pages from projects the caller cannot read under per-project authorization, and walking through them to reach others: in multi-user mode an authenticated user without a grant saw the paths, titles, kinds, and project names of pages in a restricted project up to three hops away (never their bodies). page_links and the graph already hid them; the multi-hop related walk now filters every hop the same way. Installs without authorization, and root, are unchanged. (#999)
  • Fixed CLI ai-memory message send (POST /admin/messages/send) bypassing message_send admission: it inserted the message without consulting a configured webhook's reject policy or notifying observers, while the MCP send path enforced both. It now runs admission at the recipient scope before the insert — a rejection returns 403 and stores nothing — and notifies observers only after a successful commit, without waiting on nonblocking ones. (#756)

2.5.1 - 2026-10-01

Fixed

  • Fixed ai-memory run --yolo's ai-jail re-exec aborting when the wrapped command carried a flag that ai-jail also defines: run claude --yolo --env GH_TOKEN=… failed with "flag --env after command would be passed to the child". The invocation now separates ai-jail's sandbox flags from the wrapped command with --; forwarding such a flag also needs ai-jail 2.4.2 or later, whose guard honors the separator.
  • Fixed the --yolo ai-jail offer appearing when accepting it could not work. It is now shown only on Linux/macOS when both ai-jail and its sandbox backend (bwrap / sandbox-exec) are present — never on Windows, even with a file named ai-jail on PATH — and otherwise the run proceeds without the question. The re-exec runs the exact binary that was found, so a ~/.local/bin-only ai-jail no longer fails to exec after the user accepted.
  • Fixed --true-yolo. It now implies --yolo (the warning, the ai-jail offer, and each harness's dangerous mode), so passing it alone no longer bypassed Claude's permissions with no warning; it is interchangeable with --yolo for non-Claude harnesses instead of printing "ignoring it"; and it is recognized after native arguments (run claude --model opus --true-yolo) instead of being passed to Claude as an unknown option. The claude_true_yolo config key now only upgrades an explicit --yolo/--true-yolo launch, as documented, rather than applying bypassPermissions to every managed Claude run.
  • Fixed relaunching right after an interrupted ai-memory run failing with "workstream is already active: owned by … until …" when the previous launcher could not release its lease (killed, terminal closed, or an ai-jail sandbox torn down). An interactive launch now names the holder and waits for that lease to lapse (at most one ~90-second lease; Ctrl-C aborts), then starts by itself. A holder that renews the lease meanwhile is reported as a launcher still running — never displaced — and non-interactive launches keep the short retry window.
  • Fixed --true-yolo claiming protections it never provided. It set three CLAUDE_CODE_DISABLE_*RM* environment variables that Claude Code does not read, and passed an empty permissions.ask array that cannot clear ask rules from other settings scopes (Claude Code unions them). Both were removed; true-yolo now forces only bypassPermissions, and the docs state that Claude still honors your own ask rules and command-safety checks in every mode.

Security

  • Fixed GHSA-vh98: a capture-exclusion candidate or shell argument spelled with a leading // (e.g. //repo/secret/token.txt) self-classified as a Windows UNC path regardless of the actual host, so it matched zero POSIX ignore_paths patterns (a flavor mismatch) and was captured instead of dropped. Path flavor for an untrusted candidate is now derived from the host (the cwd) rather than the candidate string alone, both in the native hook (ai-memory-hooks capture_policy.rs) and the generated OpenCode/OMP/Pi/OpenClaw TypeScript integrations (ai-memory-cli render_shared.rs); a genuine Windows/UNC host's UNC candidates are unaffected.

2.5.0 - 2026-09-30

Added

  • ai-memory run --yolo now warns before disarming an agent's safety prompts and, on Linux/macOS, offers to re-run the session inside ai-jail when it is installed. The warning and offer are TTY-gated (stdin and stderr must both be real terminals) and skipped entirely when already running inside ai-jail (detected via the Linux ai-sandbox hostname or the macOS (jail) PS1 prefix), so hook, CI, detached, and piped runs are unaffected. Accepting the ai-jail offer re-execs the original invocation under ai-jail --network --agent-state <harness> --env <NAME>... (only already-set credential/config env vars are forwarded); declining --yolo itself aborts before the agent spawns. A new opt-in --true-yolo flag / claude_true_yolo config key additionally silences Claude Code's residual --dangerously-skip-permissions prompts (the rm timeout and confirmation, and the PowerShell rm deny) and forces --settings '{"permissions":{"defaultMode":"bypassPermissions","ask":[]}}'; it is a no-op for every other harness. See docs/design-yolo-safety-ai-jail.md. (#994)

  • ai-memory repair-backfill-timestamps [--project] [--workspace] [--confirm] [--json] and POST /admin/repair-session-times correct sessions.started_at/ended_at for sessions an older backfill imported before it carried the transcript's own event time, which flattened every imported session onto the import day. Only rewrites a row whose started_at actually postdates the candidate's own transcript end (a correctly hook-captured session, or a re-run, is a no-op); refuses out-of-scope, unchanged, negative, inverted, or future-dated candidates; never closes a still-open session; and records an audit_log row on apply. Dry-run by default. (#921)

  • Native tool output resolution for Antigravity CLI lifecycle hooks (post-tool-use). Command stdout (run_command), file views (view_file), search results (find_by_name, grep_search, search_web), and task management outputs are now read directly from .system_generated/steps/<stepIdx>/output.txt and populated into observation excerpts. Code modification capture via write_to_file and replace_file_content remains preserved from invocation arguments. (#966)

  • The watcher's reconcile pass can now optionally tombstone a wiki page whose file has disappeared from disk, closing part of the gap left by #929 (the watcher only reconciled create/modify events, so a deleted file stayed indexed until ai-memory delete-page ran explicitly). Opt-in and default false ([maintenance] reconcile_tombstones_deleted_pages), so existing installs see no behavior change unless they turn it on. When enabled, a page's file must be observed missing on two consecutive 30s reconcile passes, re-verified against the live filesystem immediately before acting; a circuit breaker refuses to act on a whole scope when more than max(3, 50%) of its candidate pages look missing in one pass, or when a non-partial walk finds nothing at all (either shape is far more likely a walk/mount problem than genuine mass deletion); and reserved/indexed-but- unwalked paths (bootstrap.md, _meta.md, _pending/ sidecars, and sessions/*.md — excluded because a same-workspace move-session re-home can leave one with a correct row and no file, a separate pre-existing bug) are never candidates. The action is a soft tombstone (is_latest = 0 + superseded_at, the same shape decay eviction already uses), picked up by the SAME aged-tombstone hard-delete sweep decay eviction uses — it is not exempt from it. What actually protects it: it is never destroyed while its chain has no successor; if the file returns, the new version re-links onto the tombstoned chain instead of starting a fresh, disconnected one, so nothing is left orphaned for that sweep to destroy. It never runs the blocking admission gate (nothing can refuse it), but does fire-and-forget any non-blocking observer/mirror webhook on success. restore-page and version history remain intact. See docs/okf.md for the full design. (#964)

  • ai-memory run now finalizes the session when a harness without a native session-end hook exits (Command Code, Kiro CLI v2 and v3, Antigravity CLI), so the summary, handoff and opt-in SessionEnd consolidation no longer wait for a manual ai-memory finalize-session. Only a session the run can tie to itself (named, chosen before the spawn, or linked during the run) is closed, including again after it was resumed; if finalizing fails, the run prints the exact command to run and keeps the harness's exit code. (#941)

  • [consolidation] input_token_safety_margin (float, default 0.8, validated to (0.0, 1.0]) scales the approximate char-count input budget. The max_input_tokens budget uses a flat chars-per-token heuristic that under-budgets denser corpora — pt-BR text and source code tokenize at fewer chars per token than English and could overshoot a provider's real input limit by ~40%. The default tightens the common case modestly while leaving such corpora headroom; lower it further for a mostly non-English or code corpus. max_input_tokens is now documented as an approximate heuristic in the config reference. (#884)

  • install-skills --agent hermes (and the hermes entry in memory_install_self_routing's target_hints) installs the managed routing skills into Hermes Agent's .hermes/skills (project) or ~/.hermes/skills (global) family, completing Hermes routing support alongside its lifecycle hooks (#933). (#942)

  • [search.fts] stopwords (env AI_MEMORY_SEARCH_FTS_STOPWORDS, CSV) makes the stopword filter for bare natural-language FTS queries configurable, so a non-English install can list its own function words (e.g. Portuguese de, que, em) instead of having them match nearly every page and skew the FTS side of hybrid ranking. Absent keeps today's English list, so nothing changes on upgrade; [] disables the filter. Entries match queries case-insensitively (Unicode) but keep diacritics, so list both nao and não if users type both. Option A from #953, proposed by @rntjr. (#956)

  • docs/jev-reranker-adapter.md gains a choice-contrastive variant of the Jev reranker adapter (docs/examples/jev-reranker-adapter/jev_rerank_shim_choice.py): one choice question over the whole candidate list instead of one rubric score per candidate, choice probability mapped to relevance. The example asks for the specific page that contains the answer, and not to pick an index, catalog, or summary when a more specific page answers. On the same live golden set and 35B backend the choice shape measured hit@1 0.778 / NDCG@10 0.893 at 0.75 s mean rerank latency versus 0.636 / 0.833 at 2.2 s for the rubric adapter, and it is the shape in which replay-trained small judges hold their quality (4B: 0.808 vs 0.596). Safe because the reranker leg is sort-only; the rubric-adapter caveat on absolute relevance semantics is scoped to consumers that read absolute values.

  • ai-memory reclaim-ledger-versions drops the superseded versions of the raw hook event ledger that the pre-2.1.1 indexer left behind (#660), and nothing else. It reports what it would remove and changes nothing unless --confirm; --drop-latest also removes each ledger's live row, and --compact rebuilds the FTS index and VACUUMs to return the freed bytes. A path is only a candidate when its content is a hook ledger, so a real page named log-2026-09.md keeps its whole version chain, and only is_latest=0 AND superseded_at IS NULL rows are eligible, so decay-owned rows stay with forget-sweep. Derived FTS/entity/vector/link rows go with the page through the existing cascades, and the FTS delete trigger is stood down for the bulk delete and restored from its own sqlite_master DDL afterwards, so the cleanup does not re-tokenize tens of gigabytes of ledger body one row at a time. (#914)

  • identity = "…" in .ai-memory.toml pins the repository identity a checkout's captures route by, outranking project and the git remote — see marker-file.md. (#708)

  • The builtin web UI has a root-only /web/pending page to triage pending auto-improvement proposals. It lists the proposals of all projects, with a project filter and a sort, and shows the rationale and the proposed body. Approve and reject post to the existing /admin/pending-writes/{id}/approve|reject routes with the session cookie and the CSRF header, so admission, audit, and attribution stay the same. ai-memory-web adds no write route. (#855)

  • Native ai-memory upgrade for GitHub-release installs (Linux/macOS tarballs and Windows x86_64 zip): downloads the matching release archive, verifies the .sha256 sidecar, replaces the on-disk binary (and a sibling hooks/ tree when present), then re-runs install-hooks --apply for staged agents. Windows uses rename-aside self-replace (running .exe → .old, promote .new) because the mapped image cannot be overwritten in place. Refuses package-managed paths (Homebrew/AUR//usr), containers, and unwritable prefixes; the Docker wrapper's upgrade path is unchanged. Pin with --version / force a re-download with --force. Optional AI_MEMORY_RELEASE_BASE_URL / release_base_url overrides the Releases base for mirrors and hermetic tests (loaded via Config, not ad-hoc env). Downloads refuse bodies over 128 MiB (Content-Length and streamed cap). (#801, #802)

  • ai-memory run accepts a repeatable --env KEY=VALUE and an --env-file <path> (blank lines and # comments skipped) to pass extra environment into the spawned harness — e.g. a per-account CLAUDE_CONFIG_DIR for callers who previously had to wrap the launch in env KEY=VAL harness. Both flags are wrapper-owned like --yolo/--executable and must precede the harness name; a --env entry overrides a same-key --env-file line. The resolved environment reaches the spawned process, ai-memory's own native-session resolution, first-launch auto-wire and the global Crush config the managed context packet is layered onto, so the session store, hooks and MCP all follow a CLAUDE_CONFIG_DIR-style override, and auto-wire warns when the override puts Pi and OMP in one extensions directory. See docs/managed-workstreams.md. (#820)

  • contradiction_band_min / contradiction_band_max config keys (env: AI_MEMORY_CONTRADICTION_BAND_MIN / AI_MEMORY_CONTRADICTION_BAND_MAX) make memory_lint's A5 zero-LLM contradiction-detection cosine-similarity band configurable, defaulting to the historical fixed 0.4–0.75. The band is a fixed absolute cosine value, but background similarity is corpus-dependent: on a single-language or single-domain store, unrelated pages already sit above the general-purpose floor, so the default band measures domain proximity more than conflict and produces noisy findings. Raising contradiction_band_min trims that noise. Rejected at config load unless 0.0 <= contradiction_band_min < contradiction_band_max <= 1.0 and both are finite. (#853)

  • Fedora users can install prebuilt x86_64 and aarch64 RPMs from GitHub Releases, with the existing native systemd service assets. (#858)

  • embedding_query_prefix / embedding_document_prefix config keys (env: AI_MEMORY_EMBEDDING_QUERY_PREFIX / AI_MEMORY_EMBEDDING_DOCUMENT_PREFIX) for the openai and openai-compat embedders: an optional string prepended to query / document text before the existing truncation, so truncation still bounds the whole input. Asymmetric embedding models need a query-side instruction their publisher specifies; the OpenAI-compatible /v1/embeddings wire format has no field for it. nvidia/Nemotron-3-Embed-1B-BF16 and base E5 models use a simple "query: " / "passage: " pair; instruction-tuned E5 variants and Qwen3-Embedding instead need a task-instruction string on the query side only (documents stay plain). Empty by default — no behaviour change when unset, and not trimmed (nor is a present-but-empty env-var override, which now clears a config.toml value), so a publisher's trailing space is preserved. A non-empty embedding_document_prefix is folded into the stored embedding identity (Embedder::model_identity), so a document prefix change makes retrieval, backfill, and cleanup treat existing pages as stale and re-embed them automatically — an empty prefix keeps the pre-existing identity, so upgrading installs need no migration, and a query-only prefix change never needs a rebuild. See docs/llm-providers.md. (#859)

  • Added the optional ai-memory-relay companion for external lifecycle events. It queues events locally and sends ordered batches through the public hook API, using stable event identities for retries. The package has its own workspace and does not change the server's capture or storage defaults. (#823)

  • auto_improve.patchable_page_prefixes makes the folders whose page bodies the reviewer reads configurable, defaulting to the historical _rules/ and procedures/. Only those two folders' contents were ever sent; every other page reached the model as a single title line, so a project keeping its durable knowledge in decisions/ or gotchas/ — with no _rules/ pages at all — sent no page bodies, and the reviewer re-proposed invariants that were already written down, at high confidence, indefinitely. Existing configs are unchanged. (#834)

  • ai-memory finalize-session --reopen --session-id <uuid> re-finalizes a session that already ended, for agents without a true session-end event (Antigravity CLI, Kiro, ZCode, Pool) whose conversation continued after a first manual finalize: the discovery step now accepts include_ended=true on GET /admin/open-sessions for an exact session id (rejected without one, so reopening stays exact-id-only and never a bulk operation), and the server's normal session-end path re-runs over the new observations (updated summary page via supersession, new handoff, opt-in consolidation). Re-running with nothing new since the first end remains a harmless no-op. The default finalize behavior is unchanged: ended sessions stay invisible unless --reopen is passed. (#836)

  • Added per-execution AI_MEMORY_CAPTURE_OWNER context for external lifecycle producers. Updated native hooks, script bundles and generated integrations suppress capture while preserving supported handoff delivery and MCP recall. Documented producer provenance and stable retries through the existing hook ingestion API, without changing its schema or standalone defaults. (#821)

  • Builtin /web login and change-password pages for human password sessions: unauthenticated browser GETs to the wiki redirect to {web_slug}/login (not bare JSON 401), forms call existing POST /auth/login / /auth/password / /auth/logout, and --web-ui-dir custom SPAs stay unchanged. (#811)

  • docs/jev-reranker-adapter.md documents a stdlib-only adapter (docs/examples/jev-reranker-adapter/jev_rerank_shim.py) that serves the AI_MEMORY_RERANKER=llm request leg from a Jev /v1/systemone judge endpoint while reverse-proxying consolidation/lint/bootstrap traffic to the configured provider unchanged. In the contributor's own 102-query golden-set benchmark the judge matched the hosted reranker's hit@1/MRR/NDCG@10 (0.778/0.838/0.873 vs 0.778/0.840/0.875) at 0.205 s mean latency instead of 20.2 s — in that run the hosted mean sat on the server's 20 s completion timeout, which made the reranker stall every query before falling back. (#873)

  • MCP tool calls from OpenCode 2 (2.0.4+) are routed by the native session id the client sends in params._meta["ai.opencode/sessionID"], ahead of the Mcp-Session-Id transport header, which OpenCode shares across every session in a directory. Concurrent OpenCode sessions in different projects no longer read each other's active project; explicit workspace/project arguments still win and X-Memory-Actor-Session-Id keeps precedence. (#864)

  • OpenCode 2 turn checkpoints: every completed root turn (session.execution.succeeded, .failed or .interrupted, OpenCode 2.0.10+) now refreshes sessions/<id>.md and the session's automatic handoff without ending the native session, because OpenCode 2's shared service outlives the CLI and closing a terminal is not a session end. The checkpoint is deterministic (no LLM call), is written in the session's own scope, keeps one open baton per live session (refreshed in place, audited as refresh_handoff), and never touches a session that already ended. Child sessions neither claim startup context nor publish batons. (#865)

  • docs/llm-providers.md now has a dedicated OpenRouter subsection and a matching row in the recommended-defaults table. The wiring (openai-compat + AI_MEMORY_LLM_BASE_URL=https://openrouter.ai/api/v1) and the HTTP-Referer / X-Title app-attribution headers were already shipped, and docker/.env.production.example already ships an OpenRouter default, but the provider-facing doc mentioned OpenRouter only inside the generic openai-compat row. The new subsection covers a full working env-file, the openai-compat embedder path (with a note to verify any provider's /v1/embeddings endpoint before relying on it), a pointer to docs/llm-provider-comparison.md for model selection, and a security / gotchas block (auth-token requirement for non-loopback binds, env-file hygiene, cheap-model consolidation drift, :free-tier shared-pool limits, reasoning-model incompatibility). (#949)

  • docs/backup.md documents the remote-git-mirror backup pattern for a single-user install: what to include, what to exclude (derived SQLite index, models cache, logs, secrets), how to schedule with a systemd --user timer, how to restore, and the security posture per SECURITY.md, including a "what ends up in your wiki" section that names the exposure (sanitized prompts, tool I/O, page bodies) and lists encrypted-archive alternatives (age, restic, borg, git-crypt) for cases where a private mirror repo is not enough. A worked example ships under docs/examples/backup/ (snapshot script, .service and .timer unit files, .gitignore for the mirror repo). Pointers added from docs/deploy.md#backups, docs/airgapped-install.md, and the README docs table. The on-box ai-memory backup --to <tarball> command (docs/lifecycle-ops.md#backup) is unchanged. (#950)

  • memory_handoff_accept returns a status next to handoff: claimed when the call took the handoff, consumed_by_hook when the calling session's own SessionStart already claimed it (so it is in that session's context), and none_pending when nothing is left to claim. A bare {"handoff": null} used to cover both of the last two. consumed_by_hook needs the session id the hook claimed under, which Claude Code's install-mcp --session-aware bridge and OpenCode 2 forward; the answer stays inside the caller's project and handoff ownership, so a forwarded id cannot report another operator's or another session's claim. The handoff field is unchanged. (#988, #920)

  • Hook capture can be routed to a different ai-memory server per repository. ai-memory server add|list|remove registers named server profiles locally, in <data_dir>/servers.toml, with each token in its own owner-only file under <data_dir>/auth-tokens/. A repository's .ai-memory.toml then selects one with server = "<name>". The marker carries only the name, never a URL or token. When several profiles are registered, each must declare --root directories, and a marker outside them cannot select it. An unknown, tokenless, out-of-roots or unrooted profile, or a malformed servers.toml, makes the hook emit nothing, never fall back to the install default.

    • Native hooks route the spooled event, the session-start handoff fetch, later drains and the one-time boot backfill.
    • Generated TypeScript integrations and the .sh/.ps1 script hooks cannot route yet, so they drop a routed repository's events and handoff fetch.
    • Clients older than this release ignore the key, so upgrade every client before adding server to a shared marker.
    • Re-running server add keeps a profile's roots when --root is omitted, and changing its URL without a new token discards the old token.
    • uninstall removes the stored profile tokens together with the hook token; the registry of URLs stays.
    • hook --check-capture reports server_profile and server_resolution.

    See docs/marker-file.md. (#992)

Changed

  • ai-memory purge-session without --confirm now previews what a confirmed purge would delete before refusing, the same way purge-project does (#945): Would purge session from <ws>/<proj>: N observations, N handoffs, N pages, N auto-improve runs. (the session id itself is never printed). The server takes a new dry_run field on POST /admin/purge-session that always wins over confirm. It only counts; no delete, wiki file removal, webhook, audit row or checkpoint runs. POST /admin/purge-session's confirmed report also gains two new fields, always present on both the preview and the confirmed response: collateral_observations_deleted and collateral_handoffs_denulled, counting what the purge cascades into other projects through this session's own id. The CLI still exits non-zero without --confirm; against an older server, or if the preview times out, it prints only the existing refusal. (#957)

  • POST /admin/delete-workspace takes a dry_run field that always wins over force and previews what a confirmed delete would remove, with the same 404/409 a confirmed delete gives. It only counts: no admission call, directory removal, mirror dispatch or checkpoint. The preview and the confirmed response now also report sessions_deleted, observations_deleted, handoffs_deleted and embeddings_deleted (previously only projects, pages, workstreams and managed runs), plus collateral_observations_deleted and collateral_handoffs_denulled for rows in other workspaces reached through this workspace's sessions. merge-workspace is unaffected. There is no CLI subcommand, so this is HTTP-only. (#963)

  • Captures from a checkout with no declared project and a git remote now route by the repository's identity — the normalised upstream remote, else origin — instead of the folder name (projects.identity, V70). Two unrelated repositories sharing a folder name now get two projects (the second named after its owner, e.g. orgb-api); one repository cloned into differently named folders converges on one project, the first to claim the identity — an existing second project keeps its history but stops receiving captures. Existing projects are claimed in place, so upgrading moves no memory, and a user who may not write to a project cannot claim its identity. Checkouts that declare project in their marker route by name exactly as before. Every hook client (native, shell, PowerShell, TypeScript) resolves the identity host-side and sends it as identity / identity_src; credentials in a remote URL never leave the machine. (#708)

  • Grok Build CLI shows a pending handoff, and an opted-in [briefing], as PostToolUse additionalContext on the first tool of a session. SessionStart and UserPromptSubmit still do not accept the handoff (Grok discards that stdout); a session that never calls a tool leaves the handoff open for memory_handoff_accept. The note is clipped to 10,000 characters (Grok's own cap). Because Grok reuses one session id across a SessionEnd→restart, memory_handoff_accept now reopens an already-ended receiver session (clears ended_at) instead of rejecting it — but only for agents that reuse their session id across a restart (Grok); for every other agent an ended session stays final, so a lifecycle-only receiver still cannot reclaim after it released and ended. The reopen also only happens after the exactly-once claim guard, so a session that already took a baton still cannot take another (the multi-session claim-once invariant is preserved). Delivery of this PostToolUse handoff is exempt from AI_MEMORY_CAPTURE_OWNER capture suppression, like the other context-delivery events. (#840)

  • The generated OpenCode 2 plugin binds to the OpenCode 2.0.10+ event and hook API (session.execution.*, session.text.ended, session.moved, the context hook). Startup context is claimed once per root session and retained on every later model request; content-only tool results are captured; each location instance owns its own queue, spool state and cleanup; an explicit session.moved rebinds the live session to its new directory so its later end lands there. Checked against OpenCode 2.0.14. (#865)

  • Quieted the default server log: the reconciliation-pass summary that fired every 30 s regardless of activity dropped from info to debug, and the default log filter now pins the external rmcp MCP SDK to warn (its per-request lifecycle logging at info was the other half of a near-empty server's log). Both are restorable through log_level (e.g. "info,rmcp=info" or "debug") or RUST_LOG; the tracing_appender=warn feedback-loop guard stays non-overridable. (#894)

  • Documented Cheaper Inference as an endpoint for the existing openai-compat provider. (#981)

  • Documented API Route as an endpoint for the existing openai-compat provider. (#997)

Fixed

  • A Kiro v3 resume that falls back to the default session store drops KIRO_HOME from the child, but auto-wire still installed hooks and MCP under KIRO_HOME; it now wires the default home that resume reads. (#820)

  • Native ai-memory upgrade no longer refuses every Linux install by probing the running executable for write (Linux ETXTBSY); it only requires the parent directory to be writable for rename-based replace. (#802)

  • Parallel OpenCode 2 sessions in one directory no longer receive each other's context. Each completed turn's checkpoint retired the automatic handoffs of every other live session there, and the next session to start was handed the handoff of a session still in use (mid-turn, or seconds after its last turn). A checkpoint now spares other live sessions' handoffs; a starting session receives a live session's handoff only once that session has captured nothing for ten minutes, re-checked inside the claim; and the claim retires older handoffs of quiet sessions but not of one in use. (#883)

  • memory_query's vector stream called the generic Embedder::embed instead of embed_query on the configured embedder, so a query/document-asymmetric embedder (Google's task-typed embeddings, or the new query/document prefixes above) embedded the search query on the document side instead of the query side. Indexed writes are unchanged; only the query-side helper moves onto the query task. (#859, #861)

  • A manual memory_consolidate now reconciles the session's durable consolidation job row. The MCP handler wrote the page directly through the consolidator without touching session_consolidation_jobs, so a session whose automatic SessionEnd job had reached the terminal failed state (that the worker never re-claims) kept showing failed even though the operator had just consolidated it. After a successful, non-dry consolidate the handler flips a failed/pending/superseded row for the session to completed; a live running lease is never touched, so a concurrent automatic worker attempt is left to settle its own row. (#890)

  • Tool-family labels no longer leak into automatic handoffs and session-page titles, and the file-activity handoff warning fires again. Every closed-tool agent stores a call's title as tool file / tool non-file / … (a partition of the calls, not a tool name); these were surfacing verbatim as Tools used: tool file, tool non-file in handoffs and, for a session whose only non-prompt observation was such a call, as the page title. The same spelling meant the "session ended without a normal stop while working with files" heuristic — which only matched the bare file spelling of the minority reserved-protocol path — never fired for a real session. A shared recognizer now maps both spellings, drops the labels from the handoff tool list and the title fallback, and drives the file-activity warning from either. (#895)

  • Generated rule slugs (_rules/<slug>.md) now fold Latin diacritics to ASCII instead of turning each accented letter into a hyphen: estável slugs as estavel (was est-vel) and retenção as retencao (was reten-o). Long titles are also truncated at a word boundary (the last hyphen inside the 60-char budget) rather than mid-word. Uses the icu_normalizer NFD decomposition ai-memory-core already depends on; no new dependency. (#886)

  • Multi-page consolidation now stores page paths with a .md extension. The LLM returns a bare path for a non-rule page (decisions/smart-model-luna), and the shared path sanitizer passed it through verbatim, so the page landed extensionless and read back as a non-portable wiki path. The sanitizer now appends .md to the filename component when it is missing (idempotent, case-insensitive), fixing the non-rule consolidation, bootstrap, and auto-improve front doors at once. (#885)

  • ai-memory run's auto-wire no longer overwrites an installed session-aware Claude Code MCP bridge with the static HTTP registration. The auto-wire sentinel is keyed by client version, so the MCP step re-ran on every upgrade and replaced the ai-memory entry wholesale; a user who had run install-mcp --client claude-code --session-aware lost the bridge on the next ai-memory run after an upgrade, silently disabling [auto_scope] per_session for their MCP calls. Auto-wire now detects an existing session-aware bridge and keeps it. (#888)

  • On Windows, OpenCode 1 and 2 sessions are found again from their checkout: OpenCode records a session's directory with forward slashes (C:/Users/me/repo), so matching only the backslash checkout path found none. ai-memory doctor reported 0 local for OpenCode while it captured sessions there, and ai-memory run / show could neither discover nor list a native OpenCode session to resume. (#882)

  • Grok Build CLI hooks capture on Windows again. Grok evaluates ~/.grok/hooks/*.json commands with PowerShell, as Codex does (#515), so the double-quoted executable path in command position parsed as a string expression and every hook exited 1 with a ParserError: sessions ran with the hook installed and nothing was captured. The Windows command now carries PowerShell's & call operator for Grok too; re-run ai-memory install-hooks --agent grok --apply to rewrite an existing install. (#887)

  • The Linux/macOS Docker wrapper now keeps its native host client in ${XDG_DATA_HOME:-~/.local/share}/ai-memory/native-runner instead of ~/.cache/ai-memory/native-runner. ai-memory run auto-wires hooks whose command is that client's path (Claude Code, Codex, Kimi Code, Command Code, Kiro CLI v3, Grok, Antigravity CLI), so flushing ~/.cache left every hook of those harnesses pointing at a missing binary. The wrapper also keeps the release's hooks/ bundle beside the client, so auto-wire no longer fails with "could not locate hooks directory" for script-based harnesses on a host where install-hooks never ran. (#874)

  • Fixed pre-push installation from linked worktrees and preserved the managed block's position during reinstallation. Configured core.hooksPath overrides and ambiguous markers are rejected without replacing the existing hook. The block now keeps its shell options and SSL_CERT_FILE inside its subshell and propagates a failure explicitly, so user hook commands after it keep their own semantics and a failing test run still blocks the push. (#824)

  • A Windows service running as LocalSystem over a user-owned data directory no longer breaks the wiki git history silently. libgit2's dubious-ownership guard (CVE-2022-24765) fails every wiki commit with code=Owner when the process account does not own the repository, but the failure was WARN-only, so capture and search kept working while no wiki checkpoint was ever committed. The startup baseline checkpoint now surfaces an owner-check failure at ERROR with the remedy (run the service as the owning user), and docs/windows.md Scenario E documents running the service under a <serviceaccount>, corrects the claim that only the data directory is account-sensitive, and notes the WinSW error-1069 / stale-password gotcha for Microsoft-account / PIN / Hello users. The owner check itself is deliberately left enabled. (#872)

  • A Windows folder no longer splits into two projects. The hook router derived a project's name from the cwd after normalize_project_path_key had ASCII-lowercased the whole drive-letter/UNC path — basename included — so a session in D:\...\Default Project was captured under default project while the CLI (which keeps the raw basename) used Default Project. Because get_or_create_project matches names case-sensitively, one folder minted two projects. The router now takes the name from the raw cwd; the cache key and cwd-prefix match keep the case-folded path, so #806 handoff stickiness is unaffected. (#871)

  • Auto-improve review no longer stages a proposal whose LLM-produced page path contains a Windows-illegal character (e.g. a : copied from a conventional-commit subject). That path passed the deliberately tolerant PagePath::new and only failed later at ensure_portable when the proposal was approved, so the learning loop queued work that could not be applied. Paths are now sanitized the same way bootstrap (#847) and per-session consolidation (#848) already sanitize theirs, before validation; a path that is still unportable after sanitizing is rejected instead of staged. (#850)

  • Omitted temperature for gpt-6-* models in the codex, openai-oauth, copilot and openai providers. Codex answered GPT-6 requests that carried temperature with 400 Unsupported parameter: temperature, which broke consolidation, lint and bootstrap on a provider or fallback configured with gpt-6-luna, gpt-6-sol or gpt-6-astra. The openai provider now also sends max_completion_tokens for GPT-6 and no longer applies its local 16,384-token cap to it, matching gpt-5*. (#851)

  • memory_message_pop and memory_message_list no longer return a silent empty result when the inbox scope was inferred rather than named. A caller with no explicit workspace/project and no forwarded hook-session id resolves the shared active-project slot (whichever project published last), so two same-operator agents can have a no-scope pop land on a different inbox than the on-start notice / memory_briefing counted — "you have mail" followed by an empty fetch, with no way to tell it was the wrong inbox. An empty read from an inferred scope now reports the resolved_scope (workspace + project), the scope_source, and a hint to re-run with explicit scope; an explicitly-scoped or session-bound empty read is unchanged. No message is lost — the mis-scoped pop consumes nothing. (#854)

  • companions/ai-memory-macos/build.sh no longer fails on machines whose active developer directory is Command Line Tools only: SwiftUI @State needs the SwiftUIMacros plugin shipped with full Xcode, so the script now exports DEVELOPER_DIR to Xcode (or a caller-set path) before swift build, with a clear error when no macOS platform is present. (#849)

  • ai-memory serve no longer leaked file descriptors from half-open HTTP connections until EMFILE, breaking the healthcheck (an unauthenticated availability/DoS). A hook or MCP client whose peer died without sending FIN (laptop sleep, a VPN/Tailscale flap, an abrupt kill) left its accepted socket ESTABLISHED forever, since the OS default has TCP keepalive off — each dead peer leaked one fd, exhausting the 1024-fd default in roughly 2-3 days of normal churn. Accepted connections now get TCP keepalive via socket2, tunable with the new tcp_keepalive_secs config key (default 60s; AI_MEMORY_TCP_KEEPALIVE_SECS=0 disables keepalive). This closes the half-open-socket half of the fd leak; the rmcp session-table half was already fixed in 2.4.0 by the rmcp 2.x bump. (#792)

  • The Windows release checksum (ai-memory-windows-x86_64.zip.sha256) is now written with a LF terminator instead of CRLF. Out-File's Windows line ending made sha256sum -c fail with No such file or directory — the CR is read as part of the filename — on the WSL2 and Git Bash paths where that is the natural command, and placed a stray byte in the release body's checksum block, which concatenates every platform's file. The zip's smoke test now requires LF rather than tolerating either, so the format the release claims is the format it ships. (#838)

  • purge-session now removes every page version the session owns at sessions/<id>.md (including versions written before OKF sources existed and summaries of sessions that never recorded a summary pointer), while a manual page at the same path survives. (#862)

  • Shell hooks no longer pin a CPU core for minutes on a large payload. The hooks/_lib.sh extractors for cwd/workspacePaths/workspace_roots, the session id, and Antigravity's invocationNum located each key with ${payload#*"key"}, which is quadratic in the payload size under dash and bash: a 200 KB Cursor postToolUse event spent minutes in ai_memory_extract_cwd, and concurrent hooks stayed at 100% CPU before ever reaching the POST. A shared ai_memory_after_key helper now finds the first occurrence with one linear awk pass (about 50 ms at 200 KB) and feeds the unchanged sed parsing, so the extracted values are the same as before. (#870)

  • The generated TypeScript integrations (OpenCode 1 and 2, OMP, Pi, OpenClaw) no longer flash a console window on Windows for every captured event: their git lookups set windowsHide. The repo-root project lookup behind those spawns is memoized per cwd instead of running two synchronous git processes on every event. (#863)

  • A SessionEnd whose resolved scope drifted from its session (a .ai-memory.toml appeared under the running session) was refused as a foreign-scope end and stranded the session open forever. It now ends the session when owner and agent match and the event comes from the session's own (normalized) cwd; a different cwd, operator or agent is still refused. (#865)

  • Shell hooks on macOS no longer corrupt non-ASCII characters in the query string. /bin/sh there is bash 3.2, which sign-extends bytes >= 0x80, so ai_memory_url_encode sent é as %FFFFFFFFFFFFFFC3%FFFFFFFFFFFFFFA9 instead of %C3%A9. An accented cwd reached the server as a different path, and Cursor events and the session-start handoff lookup both use the query cwd. (#877)

  • memory_query now returns global_scope_hits (standing _global user/team preferences) for a single-project query whose project is named explicitly with workspace+project, not only when scope is omitted. The routing doctrine tells static MCP clients to pass workspace+project on every call, which set the old gate's "no named scope" condition to false, so those clients never received global preferences despite the documented contract. The union now keys on single-project resolution (scopes empty); only an explicit multi-scopes set opts out, and global=true/as_of are unaffected. The reserved-scope union is still keyed strictly to _global and never leaks another project's pages. (#930)

  • A page file rewritten under wiki/<ws>/<project>/ (the OKF import path) now gets its new version embedded the same way a brand-new file does. The watcher's reindex_page upserted the new version and stopped — embedding only ever ran on the write_page API path — so a rewrite left hybrid search silently degraded to FTS-only ranking for that page until someone ran ai-memory embed by hand. (#958)

  • [capture] ignore_paths now covers shell commands. Shell tools (Bash, shell, execute_bash, terminal, …) were classified as non-file and always kept, so cat docs/adr/*.md stored the ignored file's full text in the observation body. The native ai-memory hook now splits the command line lexically and drops the event when an argument, resolved from the event's cwd, matches an ignored pattern or is a glob that can reach one. Variables, command substitution and commands that name no path are not followed. The marker-file reference also documents excluding large tool results that Claude Code saves and re-reads from ~/.claude/projects/**/tool-results/**. (#946)

  • The generated OpenCode, OMP, Pi and OpenClaw integrations now apply the same lexical shell-command ignore_paths matching as the native hook, so a bash call such as cat docs/adr/0001.md is dropped there too instead of being captured. Both matchers now also recognize OpenClaw's and Devin's exec shell tool, resolve relative arguments from a shell tool's workdir (OpenCode bash, OpenClaw exec, Codex shell) instead of the event cwd, and treat dir/** as covering dir itself when dir holds a glob (docs/a?r/**), as the generated plugins already did. Refresh or reinstall generated plugins to pick it up. (#948)

  • Grok Build CLI tool observations are no longer stored with an empty body. Grok posts Claude Code's snake_case tool fields (tool_name / tool_input / tool_use_id), but it was missing from both closed_tool_agent and the tool-metadata agent match, so every PostToolUse body extraction returned nothing while the observation itself was still captured. Grok now shares the Claude Code tool mapping, so tool family, outcome and output land in the body. (#931)

  • ai-memory bootstrap on a repository small enough for one chunk no longer asks the provider for 64K output tokens. The output cap was keyed on the number of chunks, so the only chunk of a small repo got the one-shot cap meant for --chunk-input-tokens 0, and every such run failed on a 64K-context model. Under chunking (the default) every call now asks for up to 16K; only --chunk-input-tokens 0 keeps 64K. The --max-input-tokens help no longer claims its 150K default leaves room for 64K of output in a 200K window. (#928)

  • ai-memory bootstrap now leaves headroom for its own token estimate. It counts bytes ÷ 4, which undercounts non-English text and source code (about 40% on Portuguese mixed with code, as measured for consolidation), and it filled --max-input-tokens and --chunk-input-tokens to the last estimated token, so a chunk sized to fit a model's window could overflow it on input alone. Prunes and chunks now fill 80% of each budget by the estimate, the same default consolidation uses; a run may plan more chunks than before. (#937)

  • The bootstrap.md manifest no longer shows bare --- separators when a chunk returns no rationale. Empty rationales are dropped before the per-chunk ones are joined, and a run where no chunk returned one says so. (#939)

  • Multi-page consolidation (memory_consolidate with multi_page=true) no longer overwrites a pinned page. The batch's page paths are chosen by the model, and an update that named an existing pinned page replaced its body and wrote the new version unpinned, despite pinned pages being documented as immutable to automation. Such updates are now skipped with a warning; the rest of the batch is written. _slots/ pages, which are pinned automatically, keep their state/invariant rules. (#934)

  • Session consolidation no longer writes a page title that already exists in the project. A colliding session title gets a deterministic (session <8-char-id>) suffix (stable for the same session, distinct across sessions) and a matching leading H1 is retitled with it. The consolidator prompt tells the model to name THIS session rather than a generic harness-run phrase and not to reuse listed titles; that wording is compact enough that the advertised 6000-token input floor still projects observation bodies instead of dropping them. (#926)

  • The web page view now links a [[wikilink]] on a line indented four spaces that is not code. A nested list item written with four spaces (- Decisions: then - see [[decisions/auth]]) or a paragraph's continuation line showed the wikilink as literal text, although the engine indexed it and listed the page in the target's backlinks. The preprocessor now skips exactly the code blocks and inline code the renderer's parser reads as code. (#955)

  • observations.title is now sanitized before it is truncated, not after. title_hint used to be cut to 80 chars in ai-memory-hooks::payload before the sanitizer ever ran, so a secret straddling that cutoff was often left as a fragment too short to match a built-in or [sanitize] extra_patterns rule — landing in the title, its FTS index, and every surface that renders titles (session pages, briefings, handoffs, search) unredacted, even though the same observation's body was correctly scrubbed first. title_hint extraction now keeps the full first line untruncated; Sanitized::new scrubs the title and only then applies the 80-char display cap (ai_memory_core::sanitize::truncate_for_title), mirroring the order the body already used. (#982)

  • The /web page view keeps a leading H1 that is not the page title. It dropped the body's first H1 whatever it said, as a duplicate of the title in the header, but a frontmatter title: outranks the H1 and a setext H1 never names the page, so a heading like # Token refresh after sleep under title: Auth decisions vanished from the rendered page. An H1 that repeats the title is still dropped. (#967)

  • Shell-command ignore_paths matching no longer joins an argument vector before splitting it, which broke a path with spaces (["cat", "private notes/x.md"]) apart and let one element's stray quote hide the elements after it; each element now counts whole and is split on its own. An invalid .ai-memory.toml now makes a shell command metadata-only, like a file tool, instead of keeping its command and output, including one whose command is missing or unparseable (web_search runs nothing and is still kept), and the server now does the same when it cannot parse the client's capture marker. A long bash -lc "<script>" element is read only as words, not also as one path, so it can no longer exhaust the match budget and drop an innocuous event. Applies to the native hook and the generated plugins; the server accepts the new metadata-only shell form, so upgrade it together with them (an older server drops such an event). (#973)

  • ai-memory-importer omc-wiki now reads the frontmatter of a page saved with CRLF line endings or a UTF-8 BOM, as a wiki checked out on Windows with core.autocrlf=true is. It missed the fence, so the page's kind, tier, tags and pin were dropped and the YAML block was imported as the top of the body; the fence check now matches the wiki's own parser. (#970)

  • ai-memory serve --web-ui-dir no longer panics at startup when the custom SPA's index.html starts with a UTF-8 BOM, or has any other non-ASCII text before <head>. The <base href> injection scanned the page a byte at a time and sliced inside the multi-byte character ("byte index 1 is not a char boundary"); it now steps a whole character. (#969)

  • The /api/v1 single-page route's ETag now covers the whole JSON it returns. It hashed only the markdown body and author, so pinning a page, a frontmatter edit, or a new backlink changed the response without changing the tag, and a client revalidating with If-None-Match got 304 and kept the stale page. (#971)

  • A wikilink or markdown link written inside an inline code span is no longer indexed as a link. The engine skipped only fenced blocks, so a page showing the syntax as code (`[[other-project:notes/x]]`) got a lint broken_link finding for a dependency it does not have, and a local example listed the page in the target's backlinks, while the web page rendered neither as a link. A link whose label is code ([`foo`](foo.md)) is still indexed. (#968)

  • export-okf's generated index.md no longer has a prose sentence outside its list structure. Some strict OKF v0.2 validators read §11.3 as rejecting it. (#979)

  • install-hooks --apply --as-user <user> --auth-token <key> no longer fails with --as-user '<user>' requires --auth-token when the token was supplied. The guard was handed the rendered credential, which is deliberately None on the #552 secure path (the token is persisted under the data dir for the hooks to read), so the recommended multi-user install command in docs/users.md and install-hooks --help bailed on every native install while reporting the token as absent. It now validates the resolved token. install-hooks without --as-user was unaffected, and --as-user with no token anywhere still bails. (#993)

  • Wiki link extraction and the wikilink export now follow CommonMark for code fences and link destinations: a fence closes only on the same glyph and at least the opening length (so a ``` line inside ~~~ or a four-tick fence no longer ends it), and a backtick line whose info string holds a backtick is text, not a fence that hides the rest of the page from the link index. [doc](notes/foo_(1).md) keeps its balanced parentheses, <...> destinations and trailing titles are parsed (a ) or link inside a title is no longer read as part of the link), and a destination scan is bounded so a line of unclosed [a]( cannot stall a page write. (#985)

  • [routing] mid_session = "sticky" was silently disabled for every marker-covered install. workspace is a required .ai-memory.toml key, so the host hook forwards &workspace=… on every event under any marker's tree — including events that never left the session's own workspace — and overrides_permit_sticky (ai-memory-hooks::router) disqualified stickiness on that presence alone, before the project-provenance logic (project_src=repo-root vs. marker) ever ran. A mid-session cd into a sibling checkout therefore always rescoped under a marker, exactly as if sticky were unset, while the identical scenario outside any marker's tree (workspace_override naturally None) worked as documented — the entire difference between the two cases this bug reports as "identical except for being under $HOME". find_session_scope now runs before the sticky-permit gate, and a workspace_override is resolved once against the session's own workspace (workspace_override_is_rescope, via the existing no-create lookup_existing_workspace — no new lookup mechanism): the SAME workspace is not a rescope and falls through to the existing project-provenance logic unchanged; a genuinely DIFFERENT or unresolvable workspace still fails closed and disqualifies sticky, exactly as before. (#984)

  • Auto-improve no longer rejects every proposal from a model that spells the full-page edit mode as "full". edit_mode had the same shape #458 fixed for operation: a free-form string validated by exact match, no schema constraint, and a system prompt that says "Full-page proposals" without the literal value. gpt-oss-20b via LM Studio answered "full" for every candidate, so runs finished with zero accepted proposals and only unsupported_edit_mode rejections. The schema now advertises ["full_page", "patch"], and normalisation folds full, full-page and Full Page into full_page (with a warning) for providers without constrained decoding. Unknown modes still fail validation. (#991)

  • The one-time SessionStart backfill now delivers to the server the hook is installed against, authenticated with the persisted hook token. Before, it resolved the server from the environment and fell back to http://127.0.0.1:49374 when AI_MEMORY_SERVER_URL was unset where the agent was launched. (#992)

  • Kimi Code no longer receives a bare {} when a UserPromptSubmit is dropped by capture policy: the hook now writes an empty success response Kimi accepts instead of a JSON object it rejects. (#996)

2.4.2 - 2026-09-29

Changed

  • Documented Cheaper Inference as an endpoint for the existing openai-compat provider. (#981)
  • docs/llm-providers.md now has a dedicated OpenRouter subsection and a matching row in the recommended-defaults table. The wiring (openai-compat + AI_MEMORY_LLM_BASE_URL=https://openrouter.ai/api/v1) and the HTTP-Referer / X-Title app-attribution headers were already shipped, and docker/.env.production.example already ships an OpenRouter default, but the provider-facing doc mentioned OpenRouter only inside the generic openai-compat row. The new subsection covers a full working env-file, the openai-compat embedder path (with a note to verify any provider's /v1/embeddings endpoint before relying on it), a pointer to docs/llm-provider-comparison.md for model selection, and a security / gotchas block (auth-token requirement for non-loopback binds, env-file hygiene, cheap-model consolidation drift, :free-tier shared-pool limits, reasoning-model incompatibility). (#949)
  • docs/backup.md documents the remote-git-mirror backup pattern for a single-user install: what to include, what to exclude (derived SQLite index, models cache, logs, secrets), how to schedule with a systemd --user timer, how to restore, and the security posture per SECURITY.md, including a "what ends up in your wiki" section that names the exposure (sanitized prompts, tool I/O, page bodies) and lists encrypted-archive alternatives (age, restic, borg, git-crypt) for cases where a private mirror repo is not enough. A worked example ships under docs/examples/backup/ (snapshot script, .service and .timer unit files, .gitignore for the mirror repo). Pointers added from docs/deploy.md#backups, docs/airgapped-install.md, and the README docs table. The on-box ai-memory backup --to <tarball> command (docs/lifecycle-ops.md#backup) is unchanged. (#950)

Fixed

  • observations.title is now sanitized before it is truncated, not after. title_hint used to be cut to 80 chars in ai-memory-hooks::payload before the sanitizer ever ran, so a secret straddling that cutoff was often left as a fragment too short to match a built-in or [sanitize] extra_patterns rule — landing in the title, its FTS index, and every surface that renders titles (session pages, briefings, handoffs, search) unredacted, even though the same observation's body was correctly scrubbed first. title_hint extraction now keeps the full first line untruncated; Sanitized::new scrubs the title and only then applies the 80-char display cap (ai_memory_core::sanitize::truncate_for_title), mirroring the order the body already used. (#982)
  • The /web page view keeps a leading H1 that is not the page title. It dropped the body's first H1 whatever it said, as a duplicate of the title in the header, but a frontmatter title: outranks the H1 and a setext H1 never names the page, so a heading like # Token refresh after sleep under title: Auth decisions vanished from the rendered page. An H1 that repeats the title is still dropped. (#967)
  • memory_query now returns global_scope_hits (standing _global user/team preferences) for a single-project query whose project is named explicitly with workspace+project, not only when scope is omitted. The routing doctrine tells static MCP clients to pass workspace+project on every call, which set the old gate's "no named scope" condition to false, so those clients never received global preferences despite the documented contract. The union now keys on single-project resolution (scopes empty); only an explicit multi-scopes set opts out, and global=true/as_of are unaffected. The reserved-scope union is still keyed strictly to _global and never leaks another project's pages. (#930)
  • A page file rewritten under wiki/<ws>/<project>/ (the OKF import path) now gets its new version embedded the same way a brand-new file does. The watcher's reindex_page upserted the new version and stopped — embedding only ever ran on the write_page API path — so a rewrite left hybrid search silently degraded to FTS-only ranking for that page until someone ran ai-memory embed by hand. (#958)
  • [capture] ignore_paths now covers shell commands. Shell tools (Bash, shell, execute_bash, terminal, …) were classified as non-file and always kept, so cat docs/adr/*.md stored the ignored file's full text in the observation body. The native ai-memory hook now splits the command line lexically and drops the event when an argument, resolved from the event's cwd, matches an ignored pattern or is a glob that can reach one. Variables, command substitution and commands that name no path are not followed. The marker-file reference also documents excluding large tool results that Claude Code saves and re-reads from ~/.claude/projects/**/tool-results/**. (#946)
  • The generated OpenCode, OMP, Pi and OpenClaw integrations now apply the same lexical shell-command ignore_paths matching as the native hook, so a bash call such as cat docs/adr/0001.md is dropped there too instead of being captured. Both matchers now also recognize OpenClaw's and Devin's exec shell tool, resolve relative arguments from a shell tool's workdir (OpenCode bash, OpenClaw exec, Codex shell) instead of the event cwd, and treat dir/** as covering dir itself when dir holds a glob (docs/a?r/**), as the generated plugins already did. Refresh or reinstall generated plugins to pick it up. (#948)
  • Shell-command ignore_paths matching no longer joins an argument vector before splitting it, which broke a path with spaces (["cat", "private notes/x.md"]) apart and let one element's stray quote hide the elements after it; each element now counts whole and is split on its own. An invalid .ai-memory.toml now makes a shell command metadata-only, like a file tool, instead of keeping its command and output, including one whose command is missing or unparseable (web_search runs nothing and is still kept), and the server now does the same when it cannot parse the client's capture marker. A long bash -lc "<script>" element is read only as words, not also as one path, so it can no longer exhaust the match budget and drop an innocuous event. Applies to the native hook and the generated plugins; the server accepts the new metadata-only shell form, so upgrade it together with them (an older server drops such an event). (#973)
  • ai-memory-importer omc-wiki now reads the frontmatter of a page saved with CRLF line endings or a UTF-8 BOM, as a wiki checked out on Windows with core.autocrlf=true is. It missed the fence, so the page's kind, tier, tags and pin were dropped and the YAML block was imported as the top of the body; the fence check now matches the wiki's own parser. (#970)
  • Grok Build CLI tool observations are no longer stored with an empty body. Grok posts Claude Code's snake_case tool fields (tool_name / tool_input / tool_use_id), but it was missing from both closed_tool_agent and the tool-metadata agent match, so every PostToolUse body extraction returned nothing while the observation itself was still captured. Grok now shares the Claude Code tool mapping, so tool family, outcome and output land in the body. (#931)
  • ai-memory serve --web-ui-dir no longer panics at startup when the custom SPA's index.html starts with a UTF-8 BOM, or has any other non-ASCII text before <head>. The <base href> injection scanned the page a byte at a time and sliced inside the multi-byte character ("byte index 1 is not a char boundary"); it now steps a whole character. (#969)
  • ai-memory bootstrap on a repository small enough for one chunk no longer asks the provider for 64K output tokens. The output cap was keyed on the number of chunks, so the only chunk of a small repo got the one-shot cap meant for --chunk-input-tokens 0, and every such run failed on a 64K-context model. Under chunking (the default) every call now asks for up to 16K; only --chunk-input-tokens 0 keeps 64K. The --max-input-tokens help no longer claims its 150K default leaves room for 64K of output in a 200K window. (#928)
  • ai-memory bootstrap now leaves headroom for its own token estimate. It counts bytes ÷ 4, which undercounts non-English text and source code (about 40% on Portuguese mixed with code, as measured for consolidation), and it filled --max-input-tokens and --chunk-input-tokens to the last estimated token, so a chunk sized to fit a model's window could overflow it on input alone. Prunes and chunks now fill 80% of each budget by the estimate, the same default consolidation uses; a run may plan more chunks than before. (#937)
  • The bootstrap.md manifest no longer shows bare --- separators when a chunk returns no rationale. Empty rationales are dropped before the per-chunk ones are joined, and a run where no chunk returned one says so. (#939)
  • Multi-page consolidation (memory_consolidate with multi_page=true) no longer overwrites a pinned page. The batch's page paths are chosen by the model, and an update that named an existing pinned page replaced its body and wrote the new version unpinned, despite pinned pages being documented as immutable to automation. Such updates are now skipped with a warning; the rest of the batch is written. _slots/ pages, which are pinned automatically, keep their state/invariant rules. (#934)
  • The /api/v1 single-page route's ETag now covers the whole JSON it returns. It hashed only the markdown body and author, so pinning a page, a frontmatter edit, or a new backlink changed the response without changing the tag, and a client revalidating with If-None-Match got 304 and kept the stale page. (#971)
  • Session consolidation no longer writes a page title that already exists in the project. A colliding session title gets a deterministic (session <8-char-id>) suffix (stable for the same session, distinct across sessions) and a matching leading H1 is retitled with it. The consolidator prompt tells the model to name THIS session rather than a generic harness-run phrase and not to reuse listed titles; that wording is compact enough that the advertised 6000-token input floor still projects observation bodies instead of dropping them. (#926)
  • The web page view now links a [[wikilink]] on a line indented four spaces that is not code. A nested list item written with four spaces (- Decisions: then - see [[decisions/auth]]) or a paragraph's continuation line showed the wikilink as literal text, although the engine indexed it and listed the page in the target's backlinks. The preprocessor now skips exactly the code blocks and inline code the renderer's parser reads as code. (#955)
  • A wikilink or markdown link written inside an inline code span is no longer indexed as a link. The engine skipped only fenced blocks, so a page showing the syntax as code (`[[other-project:notes/x]]`) got a lint broken_link finding for a dependency it does not have, and a local example listed the page in the target's backlinks, while the web page rendered neither as a link. A link whose label is code ([`foo`](foo.md)) is still indexed. (#968)
  • export-okf's generated index.md no longer has a prose sentence outside its list structure. Some strict OKF v0.2 validators read §11.3 as rejecting it. (#979)
  • export-okf now backfills a title (derived, same as derive_title) and a description (from summary, else abstract) on an exported page when missing, so a generic OKF consumer sees both §4.1-recommended keys. The backfill only ever changes the bundle's copy, never the on-disk wiki file. (#979)
  • export-okf now rewrites a page's local [[wikilink]]s to bundle-relative standard Markdown links, since a generic OKF consumer has no idea what [[decisions/b.md]] means. A cross-project or cross-workspace wikilink has no Markdown equivalent and ships untouched, as literal [[...]] text. (#979)
  • sources[].author in conformed frontmatter is now process:<agent> (e.g. process:claude-code) instead of a bare agent name, matching the OKF actor grammar's process:<id> form for automated processes (#979). This changes the default sources[].author value written for every page from now on; already-written pages are not retroactively rewritten.

2.4.1 - 2026-09-25

Changed

  • Quieted the default server log: the reconciliation-pass summary that fired every 30 s regardless of activity dropped from info to debug, and the default log filter now pins the external rmcp MCP SDK to warn (its per-request lifecycle logging at info was the other half of a near-empty server's log). Both are restorable through log_level (e.g. "info,rmcp=info" or "debug") or RUST_LOG; the tracing_appender=warn feedback-loop guard stays non-overridable. (#894)
  • An unauthenticated non-loopback bind is now announced on stderr at startup independent of the log filter (a direct eprintln!, not a filterable tracing warning), so RUST_LOG=error or a container's quiet log no longer hides it. The refuse path for a non-loopback unauthenticated host bind is unchanged — this only makes the existing warning reliably visible. (#903)

Fixed

  • ai-memory restore --force no longer deletes the live wiki/ and db/ before the tarball has been read. The archive is now extracted and validated into a staging directory beside the data, the restored store is opened there so pending migrations run and the snapshot is verified, and only then are the live directories swapped out by rename (reversed if a move fails). A truncated or corrupt tarball, an entry outside the allowed layout, or a snapshot the current binary cannot open — a backup taken by a newer release, say — previously left an empty or half-extracted data dir with nothing to fall back to; it now leaves the existing data exactly as it was. (#923)
  • OMP (OpenClaw) tool calls are recorded again. OMP was missing from the closed-tool-agent set, so its tool events fell through the OpenCode-only legacy body reader and produced an empty excerpt — nothing reached session pages, handoffs, or consolidation. OMP now renders through the Pi tool schema like the other closed-tool agents (only output is kept as the excerpt; args/details stay excluded, still through the sanitizer). (#913)
  • A page whose expires_at is a bare date (2026-10-01) now gets an OKF stale_after that names the instant the TTL expires it (2026-10-01T23:59:59.999999Z) instead of the date copied verbatim. OKF v0.2 requires every timestamp to carry an explicit UTC offset, and read a bare date as the start of that day, a day earlier than ai-memory's end-of-day TTL. An RFC 3339 expires_at is still carried unchanged. Pages already written that way are repaired in place when serve starts: index row and file, with the same version row, updated_at, generated.at and body, in one wiki commit, and nothing to do on later starts. (#917)
  • ai-memory backfill stamps each imported session and observation with the transcript's own event time instead of the moment it was imported, so a transcript from weeks ago no longer looks like it just happened. An event without a valid timestamp inherits the nearest one, and the session's start and end are the earliest and latest event times. The /hook body accepts a top-level RFC 3339 occurred_at; it must be positive and no more than five minutes in the future, and anything missing, malformed or out of bounds falls back to "now" instead of failing the hook. Because an imported session now ends in the past, it can sit below the auto-improve watermark and the experience-pass anchor (no automatic review until a newer session moves them), an opt-in observation retention window can prune its older observations right after import, and the "most recently active project" fallback after a restart may not pick a project that was just backfilled. (#919)
  • The default log filter's rmcp=warn cap (#894) no longer raises rmcp above a quieter log_level. A target directive beats the global level either way, so with log_level = "error" or "off" the cap re-enabled the SDK's warnings the operator had silenced; it now only applies when log_level is louder than warn. (#896)
  • A session with no usable prompt is titled Session <id> rather than stop or session-end. Once #895 skipped tool-family labels in the title fallback, the next candidate in a real session was the kind name the router stores for an untitled lifecycle event. (#897)
  • A wiki page with CRLF line endings is parsed as having frontmatter again. markdown::parse only matched the fence lines with a bare \n, so a page a Windows editor saved, or one core.autocrlf=true checked out, was treated as body-only: reindex/the watcher indexed it without its tier, pinned, expires_at and entities (a pinned page became decay-eligible), the title came from the filename, and the one-shot OKF file pass wrote a second frontmatter block above the authored one. The parser now accepts ---\r\n fences and leaves the body's line endings untouched. (#908)
  • Rule slugs that hit the 60-character cap keep every whole word that fits. The word-boundary cut from #886 only looked for a hyphen before position 60, so a slug whose first 60 characters ended exactly on a word dropped that word, and a hyphen early in the title (a short first word before one long token) collapsed the slug to that single word. The cut now counts a hyphen at position 60 and ignores one in the first half, falling back to the hard cut at 60. (#886, #910)
  • On Windows, ai-memory run recognises an OpenCode session as belonging to the current checkout again. native_session_in_checkout (#880) compared the stored directory with the backslash cwd exactly, while OpenCode stores forward slashes (#891), so the check never matched there; it now uses the same two spellings as the other OpenCode lookups (#882). (#906)
  • A manual memory_consolidate now reconciles the session's durable consolidation job row. The MCP handler wrote the page directly through the consolidator without touching session_consolidation_jobs, so a session whose automatic SessionEnd job had reached the terminal failed state (that the worker never re-claims) kept showing failed even though the operator had just consolidated it. After a successful, non-dry consolidate the handler flips a failed/pending/superseded row for the session to completed; a live running lease is never touched, so a concurrent automatic worker attempt is left to settle its own row. (#890)
  • Tool-family labels no longer leak into automatic handoffs and session-page titles, and the file-activity handoff warning fires again. Every closed-tool agent stores a call's title as tool file / tool non-file / … (a partition of the calls, not a tool name); these were surfacing verbatim as Tools used: tool file, tool non-file in handoffs and, for a session whose only non-prompt observation was such a call, as the page title. The same spelling meant the "session ended without a normal stop while working with files" heuristic — which only matched the bare file spelling of the minority reserved-protocol path — never fired for a real session. A shared recognizer now maps both spellings, drops the labels from the handoff tool list and the title fallback, and drives the file-activity warning from either. (#895)
  • Generated rule slugs (_rules/<slug>.md) now fold Latin diacritics to ASCII instead of turning each accented letter into a hyphen: estável slugs as estavel (was est-vel) and retenção as retencao (was reten-o). Long titles are also truncated at a word boundary (the last hyphen inside the 60-char budget) rather than mid-word. Uses the icu_normalizer NFD decomposition ai-memory-core already depends on; no new dependency. (#886)
  • Multi-page consolidation now stores page paths with a .md extension. The LLM returns a bare path for a non-rule page (decisions/smart-model-luna), and the shared path sanitizer passed it through verbatim, so the page landed extensionless and read back as a non-portable wiki path. The sanitizer now appends .md to the filename component when it is missing (idempotent, case-insensitive), fixing the non-rule consolidation, bootstrap, and auto-improve front doors at once. (#885)
  • ai-memory run's auto-wire no longer overwrites an installed session-aware Claude Code MCP bridge with the static HTTP registration. The auto-wire sentinel is keyed by client version, so the MCP step re-ran on every upgrade and replaced the ai-memory entry wholesale; a user who had run install-mcp --client claude-code --session-aware lost the bridge on the next ai-memory run after an upgrade, silently disabling [auto_scope] per_session for their MCP calls. Auto-wire now detects an existing session-aware bridge and keeps it. (#888)
  • After a managed launch, ai-memory run imported the newest native session in the checkout even when a hook in the launched harness had linked the run's own session, so a concurrent launch in the same checkout could hand it another transcript. The server now records when a session is linked during a run (schema migration V67, adding managed_runs.native_session_linked_at) and reports it in the run status, and the launcher imports that session when this checkout's store holds it (a process the child starts inherits the run id; OpenCode is checked by the session's recorded directory). An older server reports no link and keeps the previous behavior. (#820)
  • On Windows, OpenCode 1 and 2 sessions are found again from their checkout: OpenCode records a session's directory with forward slashes (C:/Users/me/repo), so matching only the backslash checkout path found none. ai-memory doctor reported 0 local for OpenCode while it captured sessions there, and ai-memory run / show could neither discover nor list a native OpenCode session to resume. (#882)
  • Grok Build CLI hooks capture on Windows again. Grok evaluates ~/.grok/hooks/*.json commands with PowerShell, as Codex does (#515), so the double-quoted executable path in command position parsed as a string expression and every hook exited 1 with a ParserError: sessions ran with the hook installed and nothing was captured. The Windows command now carries PowerShell's & call operator for Grok too; re-run ai-memory install-hooks --agent grok --apply to rewrite an existing install. (#887)
  • The Linux/macOS Docker wrapper now keeps its native host client in ${XDG_DATA_HOME:-~/.local/share}/ai-memory/native-runner instead of ~/.cache/ai-memory/native-runner. ai-memory run auto-wires hooks whose command is that client's path (Claude Code, Codex, Kimi Code, Command Code, Kiro CLI v3, Grok, Antigravity CLI), so flushing ~/.cache left every hook of those harnesses pointing at a missing binary. The wrapper also keeps the release's hooks/ bundle beside the client, so auto-wire no longer fails with "could not locate hooks directory" for script-based harnesses on a host where install-hooks never ran. (#874)
  • Fixed pre-push installation from linked worktrees and preserved the managed block's position during reinstallation. Configured core.hooksPath overrides and ambiguous markers are rejected without replacing the existing hook. The block now keeps its shell options and SSL_CERT_FILE inside its subshell and propagates a failure explicitly, so user hook commands after it keep their own semantics and a failing test run still blocks the push. (#824)
  • Isolated the pre-push test process from Git's repository environment and global/system configuration so fixture commands use their own repositories. Existing installations need to run scripts/install-git-hooks.sh again. (#824)
  • memory_query now embeds the search text with embed_query rather than the generic embed() method. Google's embedder implements embed() as embed_document (RETRIEVAL_DOCUMENT), the same task type used when indexing wiki pages, so hybrid search compared a document vector to document vectors and the vector stream could not separate query from passage. Indexed writes are unchanged; only the query-side helper moves onto RETRIEVAL_QUERY. Symmetric embedders (OpenAI, Voyage, local) keep the same vector they already returned from embed(). (#861)
  • A Windows service running as LocalSystem over a user-owned data directory no longer breaks the wiki git history silently. libgit2's dubious-ownership guard (CVE-2022-24765) fails every wiki commit with code=Owner when the process account does not own the repository, but the failure was WARN-only, so capture and search kept working while no wiki checkpoint was ever committed. The startup baseline checkpoint now surfaces an owner-check failure at ERROR with the remedy (run the service as the owning user), and docs/windows.md Scenario E documents running the service under a <serviceaccount>, corrects the claim that only the data directory is account-sensitive, and notes the WinSW error-1069 / stale-password gotcha for Microsoft-account / PIN / Hello users. The owner check itself is deliberately left enabled. (#872)
  • A Windows folder no longer splits into two projects. The hook router derived a project's name from the cwd after normalize_project_path_key had ASCII-lowercased the whole drive-letter/UNC path — basename included — so a session in D:\...\Default Project was captured under default project while the CLI (which keeps the raw basename) used Default Project. Because get_or_create_project matches names case-sensitively, one folder minted two projects. The router now takes the name from the raw cwd; the cache key and cwd-prefix match keep the case-folded path, so #806 handoff stickiness is unaffected. (#871)
  • Auto-improve review no longer stages a proposal whose LLM-produced page path contains a Windows-illegal character (e.g. a : copied from a conventional-commit subject). That path passed the deliberately tolerant PagePath::new and only failed later at ensure_portable when the proposal was approved, so the learning loop queued work that could not be applied. Paths are now sanitized the same way bootstrap (#847) and per-session consolidation (#848) already sanitize theirs, before validation; a path that is still unportable after sanitizing is rejected instead of staged. (#850)
  • Omitted temperature for gpt-6-* models in the codex, openai-oauth, copilot and openai providers. Codex answered GPT-6 requests that carried temperature with 400 Unsupported parameter: temperature, which broke consolidation, lint and bootstrap on a provider or fallback configured with gpt-6-luna, gpt-6-sol or gpt-6-astra. The openai provider now also sends max_completion_tokens for GPT-6 and no longer applies its local 16,384-token cap to it, matching gpt-5*. (#851)
  • memory_message_pop and memory_message_list no longer return a silent empty result when the inbox scope was inferred rather than named. A caller with no explicit workspace/project and no forwarded hook-session id resolves the shared active-project slot (whichever project published last), so two same-operator agents can have a no-scope pop land on a different inbox than the on-start notice / memory_briefing counted — "you have mail" followed by an empty fetch, with no way to tell it was the wrong inbox. An empty read from an inferred scope now reports the resolved_scope (workspace + project), the scope_source, and a hint to re-run with explicit scope; an explicitly-scoped or session-bound empty read is unchanged. No message is lost — the mis-scoped pop consumes nothing. (#854)
  • companions/ai-memory-macos/build.sh no longer fails on machines whose active developer directory is Command Line Tools only: SwiftUI @State needs the SwiftUIMacros plugin shipped with full Xcode, so the script now exports DEVELOPER_DIR to Xcode (or a caller-set path) before swift build, with a clear error when no macOS platform is present. (#849)
  • ai-memory serve no longer leaked file descriptors from half-open HTTP connections until EMFILE, breaking the healthcheck (an unauthenticated availability/DoS). A hook or MCP client whose peer died without sending FIN (laptop sleep, a VPN/Tailscale flap, an abrupt kill) left its accepted socket ESTABLISHED forever, since the OS default has TCP keepalive off — each dead peer leaked one fd, exhausting the 1024-fd default in roughly 2-3 days of normal churn. Accepted connections now get TCP keepalive via socket2, tunable with the new tcp_keepalive_secs config key (default 60s; AI_MEMORY_TCP_KEEPALIVE_SECS=0 disables keepalive). This closes the half-open-socket half of the fd leak; the rmcp session-table half was already fixed in 2.4.0 by the rmcp 2.x bump. (#792)
  • ai-memory bootstrap no longer returns a 500 when the LLM emits a page path containing a Windows-illegal character (e.g. a : copied verbatim from a conventional-commit subject like build(sandbox): orchestrate). Such a path passed the deliberately tolerant PagePath::new and only failed later at ensure_portable inside the atomic wiki write batch, which aborted every page in the run, not just the offending one. Bad paths are now sanitized (illegal characters replaced with -, directory shape preserved) before validation, so the run and its other pages survive; a path ensure_portable still rejects after sanitizing is skipped with a warning instead of failing the batch. (#847)
  • Per-session consolidation (consolidate_session_multi) had the same Windows-illegal-path defect as ai-memory bootstrap (#847): an LLM-produced page path containing a character like : passed the deliberately tolerant PagePath::new and only failed later at ensure_portable inside the atomic wiki write batch, losing every other page from that session's consolidation run. The path is now sanitized the same way bootstrap's is, consistently across rule-routing, per-user slot placement, and the session-anchor comparison, before validation; a path ensure_portable still rejects after sanitizing is skipped with a warning instead of failing the batch. (#848)
  • The Windows release checksum (ai-memory-windows-x86_64.zip.sha256) is now written with a LF terminator instead of CRLF. Out-File's Windows line ending made sha256sum -c fail with No such file or directory — the CR is read as part of the filename — on the WSL2 and Git Bash paths where that is the natural command, and placed a stray byte in the release body's checksum block, which concatenates every platform's file. The zip's smoke test now requires LF rather than tolerating either, so the format the release claims is the format it ships. (#838)
  • purge-session now removes every page version the session owns at sessions/<id>.md (including versions written before OKF sources existed and summaries of sessions that never recorded a summary pointer), while a manual page at the same path survives. (#862)
  • Shell hooks no longer pin a CPU core for minutes on a large payload. The hooks/_lib.sh extractors for cwd/workspacePaths/workspace_roots, the session id, and Antigravity's invocationNum located each key with ${payload#*"key"}, which is quadratic in the payload size under dash and bash: a 200 KB Cursor postToolUse event spent minutes in ai_memory_extract_cwd, and concurrent hooks stayed at 100% CPU before ever reaching the POST. A shared ai_memory_after_key helper now finds the first occurrence with one linear awk pass (about 50 ms at 200 KB) and feeds the unchanged sed parsing, so the extracted values are the same as before. (#870)
  • The generated TypeScript integrations (OpenCode 1 and 2, OMP, Pi, OpenClaw) no longer flash a console window on Windows for every captured event: their git lookups set windowsHide. The repo-root project lookup behind those spawns is memoized per cwd instead of running two synchronous git processes on every event. (#863)
  • Shell hooks on macOS no longer corrupt non-ASCII characters in the query string. /bin/sh there is bash 3.2, which sign-extends bytes >= 0x80, so ai_memory_url_encode sent é as %FFFFFFFFFFFFFFC3%FFFFFFFFFFFFFFA9 instead of %C3%A9. An accented cwd reached the server as a different path, and Cursor events and the session-start handoff lookup both use the query cwd. (#877)
  • Link extraction no longer mints a permanently unresolved row from a directory target. A relations: value whose final component is empty (sessions/) had the extension appended to nothing and was stored as the literal sessions/.md; the same target in a body link or wikilink ([notes/](notes/)) stayed extension-less. No page path can match either — page paths carry .md and latest_page_id_for_link matches exactly — so both sat in links with to_page_id = NULL and were visible only as unresolved: in ai-memory status. Both routes now skip a directory target, or a stem-less .md, with the existing warning. (#915)

2.4.0 - 2026-09-21

Security

  • Bumped rmcp to 2.x (2.2.0), resolving three MCP transport advisories: GHSA-9pj6-vhgr-3mwh (unauthenticated Streamable-HTTP session-table leak / DoS), GHSA-33f5-2c5q-wgwj (missing OAuth resource-field validation), and GHSA-9g45-5xwm-f3wc (custom headers leaking to cross-origin redirect targets). Behavior-preserving: the only source change is the rmcp::model::Content → ContentBlock rename (imported under the prior name), the feature set is unchanged, and the 23-tool MCP surface is unaffected. (#794)

Added

  • macOS menu bar companion (companions/ai-memory-macos) that bundles the ai-memory binary and hooks/ tree, governs the existing LaunchAgent, and opens the built-in web UI, ai-memory status, config.toml, the data directory, and logs. Durable memory stays in ~/Library/Application Support/ai-memory; replacing the .app does not rewrite it. Documented as a README quick-start, an install.md path, a cookbook recipe, and docs/macos.md Scenario D. (#809)
  • LLM "dream" pass — cross-session rewrite/merge of cold clusters, scheduled on idle (design-memory-aging.md buckets B2/B3/B4). Where A3 collapses near-duplicate cold clusters extractively (zero-LLM, keep-token union), the dream pass hands each cold cluster to the configured provider to be rewritten into ONE coherent page. It is opt-in LLM, OFF by default, and gated on an R2 number before it may default on: it runs only when the new [dream] enabled flag is set AND a provider AND an embedder are configured — a provider-less store keeps the zero-LLM A3 path untouched (invariant #13). It never deletes a source (invariant #16): the highest-retention member is rewritten and every merged-away member is superseded with a merge-note stub pointing at it, so the full pre-merge body stays reachable via the supersession chain + git and restore-page recovers it; page_evidence (reconsolidation + b2_dream:<id>) records which members fed each merge (the hallucinated-merge guard). The rewrite routes through the existing gated apply path (preflight_admission(Consolidate) → Wiki::apply_batch, single-writer actor, invariant #2) with dry_run first (a dry run returns the plan and calls neither the LLM nor the writer), and uses JSON-schema structured output only (invariant #7). Scheduling (B3) runs the pass only after a configurable idle window with no client activity and cancels it the moment the operator returns (a cheap cancellation flag polled between clusters), bounded to a capped number of clusters per run (invariant #5). Work is ordered surprisal-first (B4): most-novel clusters — those farthest from the nearest existing page — first. Every run returns an observable DreamReport (clusters considered, merged, pages rewritten/superseded, skipped, cancelled) so a bad run is never silent. New [dream] config section; no new migration (reuses page_evidence + supersession); no new MCP tool (still 23) (#816).
  • Belief-strength confidence over the page_evidence substrate (design-memory-aging.md bucket B1 / design-hindsight-borrowings.md §3): a read-time, zero-LLM confidence derived per page version from its evidence — distinct supporting sessions (breadth, not raw count), recency of the newest sighting, and live contradicts count — bounded to [0.0, 0.95]. It is exposed inertly everywhere it helps diagnosis: memory_query(explain=true) now reports confidence and evidence_count per hit (and belief_factor when folding is on), and memory_status reports the project's evidence_rows count — none of which changes ranking. It can optionally be folded into ranking authority as one more bounded factor inside the existing [0.55, 1.50] clamp (never a new multiplier tower) via the new [retrieval] belief_authority_weight config key, which defaults to 0.0 (OFF) so upgrades rank byte-identically. The anti-entrenchment guards are baked in: breadth weighting by distinct sessions, recency shading, a hard confidence cap, and — the caller-side guard for invariant #16 — a supersession always wins regardless of evidence (a superseded version's stale evidence never boosts it, and confidence never gates whether a write or correction takes). Turning the authority factor on is gated on a positive R2 delta (retrieval-triple / QA), not yet performed (#815).
  • Zero-LLM contradiction detection surfaced through memory_lint (design-memory-aging.md bucket A5): the lint pass now flags likely-conflicting pages by cosine-similarity band. Cold knowledge pages (semantic / procedural) whose already-stored embeddings sit in the 0.4–0.75 cosine-similarity band — "same topic, but not a near-duplicate", the shape of a likely contradiction (a pair ≥ 0.75 is A3 dedup territory; < 0.4 is unrelated) — get an advisory contradiction lint finding naming both pages, with timestamp-based resolution advice (the newer page supersedes on a timestamp basis; reconcile). It is fully zero generative LLM: it reads only existing embeddings and cosine (invariant #13), so with no embedder configured — or no embeddings for the configured (provider, model, dim) triple — it is a clean no-op, not an error, and never a provider call. It runs on the user-invoked memory_lint (MCP and admin) and is advisory-only and non-destructive: it emits a finding and never deletes, edits, or supersedes a page (invariant #16), and never persists an edge (the links table's contradicts edges are body-derived and rewritten on every page write, so a programmatic edge would be silently wiped) — hence no new migration, and no new MCP tool (still 23). The scan is bounded: one embeddings load over the already-bounded cold set, capped page and finding counts, deterministic ordering (invariant #2) (#814).
  • Cold-cluster dedup of near-duplicate episodic pages (design-memory-aging.md bucket A3): the forget-sweep can now cluster near-duplicate cold episodic pages by embedding (cosine-distance DBSCAN with an adaptive k-distance eps, minPts = 2) and collapse each cluster to one survivor — the highest-retention member, its body the extractive union of the cluster's keep-tokens, so every member's durable facts survive — superseding the other members with a merge note that points at the survivor. It runs only over the bounded cold-episodic candidate set the sweep already materialises (never O(N²) over the whole corpus), is opt-in and off by default via [decay] dedup_cold_clusters (a false default), and fully zero generative LLM: it reads only already-stored embeddings, so with no embedder configured — or no embeddings for the configured (provider, model, dim) triple — it is a clean no-op, not an error. The eps is clamped to a conservative ceiling ([decay] dedup_max_eps, default cosine distance ≈ 0.15) so it errs toward NOT merging. Non-destructive and reversible: no source is ever hard-deleted — every merged-away member stays reachable via the supersession chain and git history and is recoverable with restore-page (invariant #16) — and the merge provenance is recorded in page_evidence. Every run reports its collapses in the SweepReport. Reuses existing tables: no new migration, and no new MCP tool (still 23). Ships opt-in/off; the R2 recall no-regression proof is the gate before any future default-on (#812).
  • Entropy / boilerplate pre-filter before consolidation (design-memory-aging.md bucket A4): a pure, zero-LLM Shannon-entropy + boilerplate gate that skips low-information session pages (near-empty, whitespace, single-character, or highly-repetitive boilerplate) from the cross-session experience consolidation pass before they reach the LLM prompt, the eval gate, or apply_batch. It is advisory and non-destructive — a skipped page is not consolidated, never deleted (invariant #16) — and opt-in / off by default via [auto_improve.scheduler.experience_entropy_filter] (a false default with conservative, validated thresholds tuned so a terse-but-informative note with a file path and an error code is KEPT), so an upgrade changes no consolidation output until an operator opts in. Every run surfaces the skip count in the experience report warnings. No schema change and no new MCP tool (still 23) (#812).
  • Extractive tier-down of cold episodic pages (design-memory-aging.md bucket A2): instead of evicting a cold episodic page, the forget-sweep can now compact it — keeping the L0 frontmatter abstract:, an L1 first-paragraph summary, and an L2 regex-mined keep-token set (file paths, URLs, inline-code spans, error codes, UPPER_SNAKE constants and long identifiers), and dropping the prose body. Tier-down beats eviction because the durable facts survive while the expensive, low-signal prose does not. It is opt-in and off by default via [decay] compact_cold_episodic (a false default, so an upgrade changes nothing until an operator opts in), fully zero-LLM (regex only), and reversible and non-destructive: the rewrite goes through the wiki layer, so the full pre-compaction body stays reachable in git history and the supersession chain and is recoverable with restore-page. A new V65 migration adds a nullable pages.compacted_at marker (additive ADD COLUMN, no backfill; populated lazily by the sweep from a compacted: true frontmatter mirror) so the sweep and the curator tell a deliberately-short compacted page from a cold one — a compacted page is never re-compacted, re-evicted, or re-reported as cold. Only unpinned episodic pages compact; pinned/semantic/ procedural pages are never touched. Every run reports what it compacted in the SweepReport. Ships opt-in/off; the R2 recall no-regression proof is the gate before any future default-on. No new MCP tool (still 23) (#808).
  • Per-tier retention half-life curves (design-memory-aging.md bucket A1): the forget-sweep's decay rate can now be tuned per memory tier via an opt-in [decay.half_life_days] config table, replacing the single global λ. Each key (working / episodic / semantic / procedural) is a half-life in days, converted internally to λ = ln(2) / days, so an operator can keep episodic session history longer and working-tier scratch shorter (the mcp-memory-service 365/180/90/30 shape). An omitted key falls back to the scalar [decay] lambda, so the default (no table) is byte-identical to the previous single-λ behaviour — an upgrade changes no score and mass-evicts nothing on the first post-upgrade sweep. Pure math + config: no new column, no migration, and no new MCP tool (still 23) (#807).
  • Access reinforcement on the remaining read paths (design-memory-aging.md bucket C1): memory_read_page (a direct by-path/by-query read), its include_related link-graph walk (the walked neighbours, not just the seed), and memory_explore (the pages it surfaces — rules, slots, recent, pinned, settled) now bump access_count + last_accessed_at exactly as memory_query and memory_recent already do. A page a human opens directly, or one the graph surfaces, is used and now resists decay like a search hit. Reuses the sanctioned reinforcement path: fire-and-forget on the single-writer actor, throttled to ≤1 per (page, operator) per minute, and FTS-exempt. It is strictly additive — reinforcement only raises retention scores, never blocks, never touches the response payloads, and adds no new MCP tool (still 23) (#798).
  • Reasoning tier on the LLM synthesis paths: an opt-in reasoning argument on memory_query (its answer path) and memory_explore (borrowed from Honcho's reasoning-effort ladder; targets the 2.4 line). The knob is a schema enum minimal (default) / low / medium / high / max; an unknown value is rejected. Because the provider-neutral ChatRequest carries no per-request reasoning/effort field (the provider-level reasoning_effort is fixed at construction from config), the tier maps honestly to a per-tier max-token budget scaled off each path's base budget (answer 2 000, explore 16 000): minimal = 1x, low = 1.5x, medium = 2x, high = 3x, max = 4x — a higher tier gives the model more room to reason before its output is truncated. It only tunes the answer path: reasoning is inert unless the LLM path actually runs (answer: true with a provider, or memory_explore with a provider), so the zero-LLM default path is untouched. Omitting reasoning, or passing minimal, is byte-identical to before. No new MCP tool (still 23) (#783).
  • Dialectic answer on memory_query: an opt-in, off-by-default answer argument (borrowed from Honcho's dialectic endpoint; targets the 2.4 line). When answer: true AND the server has an LLM provider configured, the query synthesizes a concise, cited natural-language answer over the top retrieved hits and attaches it as answer: { text, citations }, where citations are the page paths the answer drew from (JSON-schema structured output, grounded strictly in the retrieved snippets). When answer: true but no provider is configured, the normal hits are returned plus a short answer_unavailable note rather than an error. With answer omitted/false (the default), no LLM provider is accessed and the response is byte-identical to before, so the zero-LLM default path is untouched. Applies to the normal single-project / scopes search; global and as_of queries ignore it. Honest caveat: the feature is new and its answer quality is not yet eval-validated — treat the synthesized answer as a convenience over the same hits and still open the cited pages before acting (#782).
  • "Pin before search": memory_query gained an opt-in pin_first argument and memory_briefing now carries a bounded pinned list (default off/absent; targets the 2.4 line). Pinned pages previously earned only a small post-RRF authority bump; they were never surfaced ahead of the search, and the briefing never listed them by the pinned column. With pin_first: true, a single-project memory_query prepends the project's bounded pinned latest pages (newest first, cap 10) ahead of the fused hits, deduped by page id so a pinned page that also matches the query appears once (marked pinned: true), and re-truncates to the requested limit; scopes, global, and as_of queries ignore it. A project-scoped memory_briefing snapshot now includes a bounded pinned list of pinned latest pages (distinct from the _slots/ path-prefixed slots) so SessionStart hot-context can show standing context. Both are backed by the new ReaderPool::list_pinned_pages; default off/empty is byte-identical to the previous query ordering and briefing shape (#780).
  • memory_read_page gained an opt-in related-pages graph walk (default false; targets the 2.4 line). Passing include_related: true adds a related array of the pages reachable from the read page through the link graph — a bounded breadth-first walk that reuses the single-hop link primitive per node, following both outgoing links and incoming back-links out to related_depth hops (default 1, hard-capped at 3). Each entry carries its path/title/kind/workspace/project plus the hop depth and edge direction (link/backlink) it was reached by; the walk is cross-project aware, dedup- and cycle-safe via a global visited set, and bounded by a total-node cap. Default-off behaviour is byte-identical to the previous single-page response (no related field) (#775).
  • memory_query gained an opt-in include_superseded argument (default false; targets the 2.4 line). When set, project and explicit-scope searches also return superseded (older) page versions across the FTS/entity/vector/graph streams, each hit labelled superseded: true so callers can tell historical versions from the current one; the current version is never marked. Default-off behaviour is byte-identical to the previous latest-only retrieval, and global=true and as_of time-travel are unaffected (#773).
  • memory_status now reports which project answered: a scope object with workspace, project, and resolved_by (explicit, session, shared_slot, startup_seed, default, or default_after_mismatch). An unscoped call from a static MCP client, whose transport session id is not a lifecycle-hook session id, returned plausible counts for a project it never named, with nothing in the response to question them; resolved_by now makes that visible. The server also logs a warning whenever an unscoped MCP read is resolved by the startup seed or by the default after a session mismatch, rather than by the caller's own hook session (#757, #774).

Docs

  • Stopped recommending AI_MEMORY_LLM_MODEL=gpt-5-mini for the openai-oauth provider in docs/llm-providers.md and docs/install.md. The Codex/ChatGPT backend only accepts a small server-defined set of model ids and rejects others (including gpt-5-mini) with a deterministic 400; the docs now advise leaving the provider default (gpt-5.5) for openai-oauth/codex, keep claude-haiku-4-5 for anthropic-oauth, and qualify gpt-5-mini for copilot as unverified. (#831)
  • docs/llm-providers.md now covers the opencode LLM provider, which has shipped since 1.x but was missing from the recommended-defaults table: OPENCODE_API_KEY as the only credential, Go as the default endpoint, Zen via AI_MEMORY_LLM_BASE_URL, the built-in default model, per-catalogue model ids, and which model goes through the Responses endpoint (#763).
  • Refreshed the LongMemEval-S retrieval benchmarks on the 2.4 tree and populated the full-dataset R2 A/B (docs/benchmarks/): local embeddings add +0.149 hit@5 / +0.254 recall@10 over zero-LLM FTS, with a clean baseline-vs-baseline determinism check and no default-ranking regression vs 2.3.x (the 2.4 features are opt-in / off by default).

Fixed

  • A failed scheduled auto_improve review no longer removes its session from the queue permanently. The scheduler claims a session before reviewing it, and the candidate query excludes any session that holds a claim — but nothing ever released one, so a review that failed (a hung provider call, or a proposal the reviewer could not stage) left a claim with no run row and that session was skipped by every later tick. The state was silent: the tick reported errors=1 once and clean runs from then on, and the only exit was a hand-written DELETE. A claim now records the failure and releases, so the next tick retries it, and parks after 3 attempts with the last error kept so a deterministic failure stops costing a review every tick instead of vanishing. The tick summary counts parked separately from errors. (#833)
  • The auto-improve reviewer now excludes sessions/ pages from its own recent-page context so those slots go to durable pages (decisions/, gotchas/, _rules/, …) it might otherwise re-propose. Session pages are never valid proposal targets and previously dominated the recency-ordered list, crowding durable knowledge out of the reviewer's view. The exclusion is scoped to the reviewer only — the SessionStart briefing and memory_briefing still include session pages. docs/auto-improvement-loop.md now documents that only _rules//procedures/ page bodies reach the reviewer and that the recent-page list is recency-ordered, with configurable patchable prefixes and embedding-nearest dedup noted as deferred future work. (#834)
  • Auto-improve proposal staging no longer discards an entire run when one proposal is a create/update misclassification. A Create whose target page already exists, or an Update/patch whose target is missing, previously aborted the staging transaction, dropping every sibling proposal and the run row over one probabilistic LLM mislabel. Those two cases now skip just the offending proposal (reported as skipped, like a pending-target collision) and keep the rest of the run. Two proposals in one run targeting the same path remain a hard error, and a create-on-existing is never coerced to an update (the page could be pinned). (#832)
  • The Windows Docker wrapper (bin/ai-memory.ps1) now forwards the same provider credentials and host-config env vars as the POSIX wrapper into the helper container. A host-exported GEMINI_API_KEY / GOOGLE_API_KEY, Copilot token, OPENCODE_API_KEY, CLAUDE_CONFIG_DIR, or AI_MEMORY_WORKSTREAM_ID previously never reached Config::load, so llm-test, Copilot auth, OpenCode, and a relocated Claude config all reported "not configured" on native Windows Docker Desktop even though the same export worked through bin/ai-memory. The POSIX wrapper also now forwards OPENCODE_API_KEY. (#803)
  • Config::load now treats Windows %USERPROFILE% (then dirs::home_dir) as the operator home when AI_MEMORY_HOME and $HOME are unset. Native Windows often has no HOME, so the #103 catch-all guard — skip a stored repo_path equal to the user profile so it cannot prefix-match every project beneath it — was inert there and a home-directory project could swallow unrelated cwds. (#804)
  • Automatic handoff selection and cwd-prefix project matching now treat Windows drive-letter and UNC paths as case-insensitive. A Linux server (the Docker Desktop helper) comparing host cwds from Explorer, Git, and PowerShell previously required a byte-exact match, so C:\Users\…\repo vs c:\users\…\repo missed the pending auto-handoff and could mint a fragment project. Unix paths stay case-sensitive. (#806)
  • Fixed Ctrl+C at the native-session chooser leaving the launcher alive and renewing its workstream lease. Cancelling now releases the acquired run and exits without waiting for Enter or linking a native session (#795).
  • The privacy strip now redacts Windows credential paths (C:\Users\…\.ssh, .aws, .kube, .gnupg, .config\gcloud). The previous patterns required a POSIX / separator, so a captured tool result that echoed a native Windows path stored the profile directory and key file name verbatim. (#805)
  • The privacy strip now redacts secrets written in JSON. The quote before a value put it outside the value character class, so {"db_password":"..."} was stored verbatim while the identical YAML form was redacted, and JSON is the shape most captured tool payloads arrive in. The same rule now also accepts an auth scheme word before the value, so Authorization: Basic <base64> (which carries user:password) is redacted like the Bearer form already was, and covers two unprefixed names the generic env rule missed: Azure AccountKey= and npm _authToken=. (#800)
  • Terminal escape sequences, NUL and bidi override characters are stripped from captured text instead of being stored. Page bodies and observations are replayed to a terminal by ai-memory read-page and ai-memory search, where an escape rewrites the screen or the window title and a bidi override reverses what the reader sees; a NUL additionally made the markdown file binary, costing it grep and git diffs. Tabs, newlines and carriage returns are kept. (#800)
  • A page write is refused when another live page in the same project differs from it only by case or Unicode normalization. Such a pair is one file on macOS (APFS) and Windows (NTFS), so creating the second silently overwrote the first page's file while the index kept both rows, so reads for either path then returned the survivor's body, and the wiki watcher superseded the overwritten row, losing the original content from disk and index alike. The refusal names both paths, applies on every platform (the wiki is synced between them), and leaves supersedes of an existing path untouched. reindex skips such a pair instead of failing the whole rebuild, reports the count, and logs each one. (#799)

2.3.2 - 2026-09-20

Changed

  • memory_consolidate accepts an omitted session_id. Omitting the field (or sending null) no longer fails deserialization with missing field session_id; the tool consolidates the latest completed session in the resolved project — the same default memory_auto_improve and memory_read_session_observations already use. Pass an explicit UUID to target a specific session, and dry_run=true for the cheap admission preflight. A project with no completed session now fails as no completed session in <scope> instead of a deserialization error.
  • A consolidation LLM call that fails on a transient provider error (429, any 5xx, a transport timeout or connect failure) is retried twice, two seconds apart, before the failure is reported — the same bounded policy bootstrap already applies to its chunks. Deterministic failures (auth, schema, a malformed-request 4xx, unparseable or truncated output) are still reported on the first attempt, since retrying them only burns another call.

Fixed

  • CLI commands no longer fail at startup when an [[llm_fallbacks]] profile's api_key_env variable is absent from the invoking shell. Config::load validated every fallback credential eagerly, so read-only commands such as ai-memory status exited with llm_fallbacks[0].api_key_env=... is set but the environment variable is missing or empty even when the running server had the key injected by its service wrapper, which pushed operators to export provider keys in every shell. The missing credential is now enforced where it is needed: ai-memory serve still refuses to start without it, and building the LLM chain still fails rather than silently dropping the fallback. Every other profile check (provider, model, base URL) still runs at load for every command (#762).

  • backfill --dry-run recorded a completed attempt and suppressed the next automatic import. Planning now leaves the backfill sentinel untouched, even for populated projects or an opted-out automatic invocation (#785).

  • ai-memory serve no longer hard-fails to take its single-instance lock on a transient error under load. Acquiring the serve lock now retries open and try_lock_exclusive a few times with a short (~25ms) backoff when they hit a transient failure (EMFILE/ENFILE fd exhaustion, EINTR), mirroring acquire_drain_lock. A genuinely contended lock (WouldBlock, another server holds it) is never retried and still refuses startup immediately. The serve-lock tests also assert with the concrete errno so any remaining environmental flake is diagnosable rather than silent (#745).

  • GitHub Copilot completion requests now select the model-advertised API endpoint from /models: existing Chat Completions remains preferred when available, while Responses-only models use /responses. Responses requests preserve strict JSON Schema structured-output constraints and report empty, refused, or rejected output without silently downgrading the contract. A model the /models catalogue does not enumerate (enterprise/custom deployments, aliases, a model newer than the list) falls back to Chat Completions with a warning instead of erroring, matching the graceful fallback already used when /models is unavailable. As a related behavior change, a Copilot chat completion that returns empty content now reports UnexpectedShape rather than yielding an empty string. (#761)

  • The V62 page-ingestion-window migration no longer runs its backfill inside a single migration transaction, which on a large store ran for hours and grew the WAL to roughly the size of the database with no progress. V62 is now DDL-only (the two columns plus their index); the window backfill moved to a chunked, resumable, WAL-bounded boot-path step that processes pages in bounded batches, checkpoints the WAL (TRUNCATE) between each, and logs progress. The end state is byte-identical to the original V62, the step resumes rather than restarts if interrupted, and it is a fast no-op on a store that already applied the original V62. Because that reshape changes the migration's checksum, the runner now intentionally tolerates a divergent checksum on an already-applied migration (abort_divergent = false) so correctly-migrated stores still open; the schema-ahead guard (abort_missing) is unchanged (#776).

  • Page writes now refuse git-reserved and non-portable page paths (a .git component or an 8.3 git~1..git~4 alias, Windows-reserved names and characters) on every write funnel, including MCP memory_write_page and consolidation apply_batch; reads of already-stored pages stay tolerant so a bad row never breaks a listing. The git-reserved check is byte-safe and no longer panics on a 5-byte multibyte path component (#781).

  • The generated OpenCode and OpenCode 2 plugins now forward a subagent session's parentID as the agent_id marker, so [capture] drop_subagent_captures can recognize and drop OpenCode subagent sessions. Previously both plugins emitted only title/projectID on session.created, so the marker never reached the server and the opt-in was a silent no-op for OpenCode. Root sessions (no parentID) stay unmarked (#755).

  • Scope-resolution failures over MCP now answer with invalid params (-32602) instead of an opaque internal error (-32603), the same split the web route applies with its 400/404: a malformed scope argument, or a workspace/project name that does not resolve, is caller input, while a missing writer handle or an underlying store failure stays internal. The messages are unchanged.

  • memory_consolidate treats a blank session_id ("" or whitespace) exactly like an omitted one — the resolved project's latest completed session — and a malformed id now fails as invalid params, the code memory_auto_improve already uses for the same argument.

  • backfill returned success even when imports failed, and --quiet hid their diagnostics. It now reports errors on stderr, includes failure counts in the human summary, and exits nonzero after emitting its report (#786).

2.3.1 - 2026-09-17

Added

  • DATA_HANDLING.md, docs/sso.md, and docs/airgapped-install.md, consolidating existing data-flow, OIDC, and offline-install facts from SECURITY.md/docs/install.md/docs/local-embeddings.md into the shape enterprise security/legal review typically asks for before approving a dev tool. No behavior change; a small addendum to SECURITY.md's reporting section adds a fallback contact path for reporters who can't use GitHub's private-advisory flow.

Fixed

  • export-okf no longer refuses to export any project that has captured an observation. The bundle walk skipped the reserved names index.md and log.md, but not the rotated hook ledger log-YYYY-MM.md the server itself appends to — which carries no frontmatter, so okf::is_conformant failed it and aborted the whole export with 422 page log-YYYY-MM.md is not OKF-conformant; run the server once to migrate before exporting. Migrating could not help: the OKF migration deliberately skips ledgers (#669), so the file the export demanded be conformed stays frontmatter-less forever. The ledger is raw capture rather than a concept file, so it is now dropped from the bundle exactly as log.md already was, sharing the migration scan's content gate — a prose page that happens to be named log-2026-09.md still ships and still has to declare a type. (#748)

2.3.0 - 2026-09-16

Added

  • ai-memory run <harness> now auto-installs that harness's ai-memory hooks and MCP server the first time it launches the harness, if they are not already wired — so managed launch (the recommended way to start a harness) captures and can query memory without a separate install-hooks / install-mcp step. It is idempotent and one-time per harness + binary version (a per-agent sentinel under <data_dir>/autowire-state/), preserves any unrelated user config the installers touch, runs before the child spawns so the harness picks up the fresh hooks, and is best-effort (a failure warns and the launch still proceeds). Harnesses without installer support (e.g. Crush) are skipped cleanly, and Pi wires hooks but has no MCP client to write. Opt out with ai-memory run --no-autowire, AI_MEMORY_RUN_AUTOWIRE=false, or run_autowire = false. Documented as the preferred launch path ("if in doubt, run with ai-memory").
  • Boot-time backfill of pre-hook local history, on by default. When a project's ai-memory store is brand new (empty), the SessionStart hook triggers a one-time, bounded import of that project's existing local harness transcripts so installing hooks mid-project no longer starts amnesiac about the very session you are resuming. It replays each local transcript through /hook — the same ingress live capture uses — so the history becomes real sessions and observations that consolidate into pages and are searchable via memory_query, sanitized and bounded on the server exactly like live capture and attributed to the original harness. It only ever bootstraps an empty project (never overwrites an established one; live capture from install-time forward and backfill of before-install history do not overlap), is hard-capped (newest 25 sessions, 50k events), and runs detached so session start is never blocked. New ai-memory backfill subcommand runs it by hand (--dry-run, --force, --session, --json). Opt out with AI_MEMORY_BACKFILL_ON_START=false / backfill_on_start = false.
  • New ai-memory doctor command: a capture-coverage check that, for the current project, compares every known harness's local (on-disk) native session store against what the server actually captured (GET /admin/sessions/by-agent) and warns when a harness ran here recently but has zero captured sessions — the silent "its hook was never installed" gap — printing the exact install-hooks --agent <name> --apply to fix it. Read-only on the local side; supports --json, --since-days, and explicit --workspace/--project.
  • New task-oriented docs/cookbook.md cheat sheet ("I want to do X" → how: recall, durable rules, importing a knowledge base and reading a specific document, two agents working together), linked from the README docs table, to make it clearer what ai-memory does and how to use it (#726).
  • Codex assistant-final-turn capture: install-hooks --agent codex --capture-assistant now captures the assistant's final message on Stop, the same double opt-in (client flag + server capture_assistant = true) and sanitize/bound pipeline as Claude Code. Codex's Stop payload carries last_assistant_message (verified on codex-cli 0.154.0); the installer previously refused the flag for Codex and now bakes it on a native hook platform. Only (Codex, Stop) is added — Codex has no SubagentStop (#743).
  • Cross-project agent messaging: a directed, claim-once inbox/queue so an agent in one project can hand a self-contained request to an agent in another project without pulling that project's context into its own session. Four new MCP tools (memory_message_send / memory_message_list / memory_message_pop / memory_message_cancel, bringing the surface to 23) plus ai-memory message send|list|pop|cancel CLI subcommands and /admin/messages* routes, backed by the new agent_messages table (migration V64). Messages are addressed to a project (any session there can pop; claim-once); the sender can retract a pending message; an unknown recipient fails closed; each inbox is depth-capped. A popped message is treated as untrusted cross-project input — secret-scrubbed and size-capped on send, fenced with a security notice and sender provenance on pop, and never auto-injected into context. See docs/agent-messaging.md.
  • The session-start "hot context" block now appends a non-consuming inbox notice (a static count only — never message text) when a project has pending cross-project mail, and memory_briefing reports pending_message_count.
  • Added an independent codex LLM provider that follows the Codex CLI account selected by CODEX_HOME, reloads its read-only auth.json credentials before each operation, and delegates expired-token recovery to codex app-server --stdio (#716).
  • install-mcp --flavor <moonshot|bedrock|gemini> pins the tool-schema dialect in the URL written into a client's config. The installer already picks one for the clients whose upstream is fixed — Kimi Code is always Moonshot, Kiro always Bedrock — but a client that fronts several models cannot be pinned by its name alone, and there was no installer path to ?flavor=gemini for any of them. A Command Code or OpenCode install routed to Vertex needed the marker added by hand or every model call 400'd at tools/list; the same client on a non-Google model does not, which is why this is an explicit choice rather than another per-client default. vertex is accepted as an alias. uninstall now matches every marker on every client, so an entry pinned this way is still removed (#735).
  • AI_MEMORY_EMBEDDING_PROVIDER=copilot adds GitHub Copilot as an embedding provider, reusing the copilot LLM provider's OAuth login and GitHub-token exchange (no separate API key). Defaults to text-embedding-3-small, 1536-dim; calls Copilot's /embeddings endpoint following the OpenAI-compatible contract Copilot documents for chat (#739).

Fixed

  • install-hooks --apply now preserves an existing --capture-assistant opt-in on a bare re-apply (one with no --capture-assistant flag). Previously a refresh without the flag silently stripped assistant capture; this most affected the new ai-memory run auto-wire, which always re-applies without the flag. The flag still explicitly enables it; an unset flag now keeps whatever is already installed (Claude Code and Codex).
  • install-hooks --apply no longer aborts the entire install when the hook bearer token cannot be persisted under the data dir. This bit the docker wrapper, where data_dir is /data — a container volume the host hooks never read from and that the container user frequently cannot write — so the persist step failed with Permission denied and left the operator with no hooks and no capture at all. The installer now falls back to embedding the credential inline in the rendered hook config (the pre-#552 behavior), warns that the token is then readable in that file and how to restore the secure on-disk path (a writable host data dir, e.g. AI_MEMORY_DATA_DIR=$HOME/.local/share/ai-memory for the docker wrapper, or a native binary), and completes the install so capture keeps working.

2.2.2 - 2026-09-15

Security

  • Updated rustls 0.23.40 → 0.23.45 for RUSTSEC-2026-0285, in which a TLS 1.3 handshake message that follows a key-changing message in the same record can be accepted at the wrong encryption level. rustls is a direct dependency — it installs the process-wide crypto provider the MCP bridge needs for HTTPS — and is also the TLS implementation every outbound HTTPS call resolves to through reqwest, so the advisory failed cargo audit and cargo deny check on every open pull request that inherited main's lockfile, including ones that change no Rust at all. Cargo.toml already allows compatible 0.23 patch releases, so this is a lockfile-only change needing no manifest or public-surface edit; the same resolution moves rustls-webpki 0.103.13 → 0.103.15 and nothing else (#731).

Fixed

  • The native ai-memory hook session-id state file (<data_dir>/hook-state/<agent>-session-id), introduced for Devin in #178, now also covers ZCode. ZCode's hook payloads do not reliably carry a session id, it fires Stop at the end of every turn, and it has no SessionEnd event — without a persisted id, each turn with an id-less payload opened a fresh server-side session that nothing ever closed (observed in production: 4 sessions with NULL ended_at and 398 observations stuck in episodic). ZCode events without a native id now share one stable stored id per agent session; events carrying a native id are passed through untouched. The stored id is cleared by finalize-session --agent zcode (which now also removes the state file) or overwritten by the next session-start; Stop never clears it. Other agents are unaffected: payloads that carry a session id short-circuit the state file exactly as before.
  • memory_feedback's signal (FeedbackKind) JSON schema now declares a top-level type: "string", so strict function-calling gateways (Moonshot/Kimi and other schema validators) accept the tools/list surface instead of rejecting the enum for a missing type (#735, #741).
  • Cursor lifecycle events are no longer stored twice on a host where both install-hooks --agent cursor and install-hooks --agent claude-code are applied. Cursor also runs the commands in Claude Code's settings, and since 2.1.0 that copy is re-attributed to cursor by its cursor_version, so it landed in the same session as the native event. The --agent claude-code hook now drops a Cursor-marked payload when ~/.cursor/hooks.json already registers an ai-memory --agent cursor hook; without Cursor's own hooks the Claude Code path keeps capturing Cursor sessions as before (#721).
  • install-hooks --agent cursor --apply now warns about existing ~/.cursor/hooks.json entries that mention ai-memory but are not ai-memory hook entries, such as a pre-2.1.0 shim that injected cwd. Those entries are kept beside the native ones, so every event would otherwise be captured twice without any sign (#721).
  • The POSIX shell hook bundle's spool reader no longer builds the decoded value in memory. ai_memory_json_field appended to a string that grows to the whole value, so reading a multi-megabyte entry — one large tool result is enough — took minutes of CPU, and a drain pass reads every entry three times (url, body, token) while one detached pass starts behind every delivery that succeeds. Once an outage had filled the spool, the passes accumulated faster than they retired and saturated the machine. The value is now bounded by a single regex pass over the JSON string grammar, unescaped with gsub over whole segments and written straight to stdout, for identical output and exit codes. A 2.2 MB entry of grep output — the shape that caused the incident, where every literal backslash is an escaped pair — goes from 732 s to 5.5 s under the awk macOS ships and from 9-10 s to about 1 s under mawk and gawk (#727).
  • The POSIX shell hook bundle now assigns an ingest_key before its initial delivery and preserves that key when spooling the event, preventing a replay from creating a duplicate observation when the server committed the first request but its response was lost (#729).
  • install-hooks reapply is idempotent again on Windows for agents whose native command uses an underscore executable name (ai_memory): the hook ownership predicate recognized only the hyphenated ai-memory, so a reapply reported Updated and failed to dedup its own prior Antigravity/Cursor/Kimi Code entries. It now matches both forms while still requiring the full hook --event … --agent … --server-url … argv signature (#740).
  • The POSIX shell hook bundle's ai_memory_json_string now escapes every JSON control character (U+0000–U+001F) as \u00XX, not just backslash, quote, tab and CR. A replayed tool result carrying an ANSI colour escape (0x1b) reached stdout bare, so the managed-workstream SessionStart packet — and any handoff whose summary held a control byte — was rejected as invalid JSON and the resuming session started with no context. The escaping is linear and reuses the same BusyBox replacement-doubling probe as the four existing escapes (#732).

2.2.1 - 2026-09-12

Fixed

  • The POSIX shell hook bundle no longer drops a lifecycle event when the server is unreachable or answers 5xx. A failed delivery is written to <data_dir>/hook-spool/ in the same on-disk contract ai-memory hook-drain reads (same filename shape, same SpoolEntry JSON, same 0600/0700 modes, tmp+rename), with an ingest_key minted once at spool time so a shell drain and a concurrent hook-drain cannot double-ingest; the backlog is flushed behind the next delivery that succeeds, detached from the hook so the agent never waits. A 4xx stays a permanent rejection and is not retried. This is the durability the generated TypeScript integrations got in #580, for the path the docker deploy installs. The PowerShell bundle is unchanged (#719).
  • The POSIX shell hook bundle's POST /hook now hard-timeouts at 200 ms rather than 500 ms, which is the budget invariant 5 documents for a script hook. A real loopback round trip runs about 0.3 ms, so a local install is unaffected; against a remote server the tighter ceiling is safe only because a missed window now spools instead of dropping (above). The handoff GET keeps its 1 s: it is fed synchronously to the agent's context, and truncating an almost-ready handoff costs more than it saves (#719).
  • ai-memory restore no longer rejects the GNU-sparse SQLite entry that ai-memory backup itself produces. tar::Builder's default sparse detection archives db/memory.sqlite as a GNU-sparse entry (header type S) whenever the SQLite snapshot has real holes on disk, which validate_restore_entry's is_file()/is_dir() check rejected as an "unsupported entry type" — a verified, valid backup could not be restored, with no recovery path but hand-editing the archive. tar already expands GNU-sparse blocks to their full logical content while iterating entries, so restore now accepts the type and unpacks it exactly like a regular file (#718).

2.2.0 - 2026-09-12

Added

  • Retrieval explain now names the typed edge a graph-stream neighbour was reached by: memory_query(explain=true) reports graph_via.edge = causes / fixes / contradicts (omitted for a plain references link), so it is visible why a page surfaced through the link graph. Retrieval ranking is unchanged — this is explanation only; typed-edge weighting and contradicts capping are deferred behind the eval harness (docs/design-hindsight-borrowings.md P3).
  • Page-grain ingestion windows (pages.valid_from / valid_to, V62) and a second as_of stream: memory_query(as_of=T) now fuses the entity timeline with version-filtered full-text search over the page versions alive at T, using the default path's RRF (k=60) plus the same bounded authority adjustment — current-index relevance over knowledge valid at T. This answers entity-less pages and paraphrased audit questions ("which database were we on during the outage?") that the entity-only lookup missed. explain=true reports both streams (streams_active: ["entity", "fts"]) with per-stream ranks and RRF contributions. Vector, graph, and the raw-observation fallback stay out of audit mode; the default (no as_of) path is unchanged. Every retire path (supersede, decay, reorg graveyard, move-regenerate) closes both grains in the same transaction, and the V62 backfill closes successor-less retirements at the decay marker, existing entity-link close, or updated_at fallback, preserving recorded reorg/move history. Phase B world-time stays deferred. (#656)
  • [retrieval] section with two opt-in ranking signals, both off by default so unconfigured stores rank exactly as before. query_intent routes lexically session-recall queries ("上次 / 之前那次…的会话 / last time / yesterday …") past the default session-page authority penalty (×0.77 combined kind/tier), with session_recall_bonus (default 0.25) sizing the lift; abstract_vectors adds a fifth RRF stream over the new page_abstract_embeddings table (V61) — the L0 layer: a page's frontmatter abstract: line is embedded on its own, at write time when an embedder is attached and by the same backfill as the body otherwise (a one-line summary embeds far more sharply than a multi-thousand-character body). Measured on a 138-query golden set over a production two-year wiki (FTS5 + entity + vector + graph, mis-tei Qwen3-Embedding-8B): hit@1 0.609 → 0.746 (+22%), NDCG@10 0.782 → 0.879 (+12%) with both enabled; every unrouted category keeps its exact baseline ordering. Cross-checked on an independently built 99-query golden set against the same store: hit@1 0.556 → 0.626, NDCG@10 0.712 → 0.779 (+9.4%). Both signals are visible per hit in memory_query(explain=true) (intent, intent_boost, abstract_rank, rrf.abstract). (#672)
  • memory_handoff_list lets no-stdout and MCP-only clients inspect open project handoffs without claiming or expiring them, then memory_handoff_accept can claim that exact handoff_id once. Grok, Zero, and other clients that discard SessionStart stdout no longer have to recover a baton by blindly accepting the latest row. Listing is owner-filtered like accept (own plus shared; root-only any_owner); it is not a second claim path, and Grok SessionStart still does not fetch /handoff. (#664)
  • install-instructions --compact writes a slimmer managed routing block for projects that already have the detailed ai-memory Agent Skills installed. The compact block keeps the same start/end markers (so refresh/uninstall still find it), the untrusted-history security scaffold, and the cross-harness memory-of-record guidance; full_block and the default remain unchanged. (#675)
  • memory_install_self_routing accepts compact: Option<bool> so agent-driven refreshes of a compact-installed file preserve the compact routing block instead of rewriting it back to full. (#685)
  • The openai-compat provider now sends OpenRouter's app-attribution headers (HTTP-Referer, X-Title) by default when its base URL points at openrouter.ai, so ai-memory's usage shows up on OpenRouter's app leaderboard. An explicit AI_MEMORY_LLM_HEADERS entry for either header still wins, and a non-OpenRouter compat endpoint (Ollama, vLLM, LM Studio) never receives them. (#686)
  • ai-memory run now accepts any claude*-prefixed harness name (claude-corp, claude-personal, ...), all resolving to the same ManagedHarness::Claude/AgentKind::ClaudeCode — no new session store, migration, or agent kind. This is for callers juggling more than one Claude account (e.g. Corporate and Personal): name each account's PATH wrapper script claude-<account> and pair it with --executable claude-<account> (bare names already resolve through PATH) so ai-memory run claude-corp --executable claude-corp and ai-memory run claude-personal --executable claude-personal each launch the right binary while reading clearly in shell history (#687).
  • install-mcp --client muse registers ai-memory with Meta's Muse Code, merging a native streamable-HTTP entry with bearer headers into the snake_case mcp_servers map of ~/.config/muse/settings.json and preserving sibling servers. The writer adds the mandatory "schema_version": 1 when it is absent — without that key every muse command fails at startup with malformed settings file — and never rewrites an existing value, so a future schema is not downgraded. The entry sets "mode": "optional" because Muse defaults it to required, which aborts the whole run when the memory server is unreachable. MCP-only: Muse documents a lifecycle hook surface, but the output contract of its SessionStart event is not specified, so capture and managed workstreams are not claimed. Skills need no extra step — Muse reads ~/.agents/skills, which install-skills already writes (#659).
  • Belief-strength evidence substrate (P2, docs/design-hindsight-borrowings.md §3): a new append-only page_evidence table (V63, (page_id, source_kind, source_id, created_at), source_kind one of session/observation/feedback/reconsolidation) records what produced or reaffirmed each page version, written in the same transaction as the page upsert. Rule-based and zero-LLM — the consolidator cites the session(s) it drew on for both the single-page and batch write paths, so the substrate populates on the default path with no provider configured. hybrid_search_explained (and therefore memory_query(explain=true)) now reports evidence_count per hit, batch-fetched once after fusion; the default (non-explained) path and ranking are unchanged — evidence is inert data and an explain field this release, not a ranking input. The confidence-into-PageAuthority step described in the design doc is deferred behind the planned R2 retrieval eval.
  • Standing-answer boot surfacing (P4, docs/design-hindsight-borrowings.md §5): memory_briefing accepts an opt-in settled_first: bool (default false, unchanged briefing shape). When true, the snapshot's new settled array leads with the project's highest-standing rule/decision pages — up to 8, ordered by page_evidence count (P2) then recency — so an agent can boot from settled answers instead of re-deriving them. Pure SQL, no LLM call; every other briefing field, and the default settled_first: false path, are byte-for-byte unchanged.

Changed

  • The privacy sanitizer now redacts to a typed marker — [REDACTED:<kind>] (e.g. [REDACTED:github_token], [REDACTED:jwt], [REDACTED:env_secret], [REDACTED:custom] for operator patterns) — instead of a bare [REDACTED], so a later reader knows what kind of secret was present without it leaking. Redaction stays complete and idempotent; the trust boundary is unchanged. Note: this changes the durable sanitized string, so pages written before the upgrade keep [REDACTED] while new writes carry the label (see docs/design-hindsight-borrowings.md P1).

Fixed

  • Preserved recorded entity-link retirement timestamps when backfilling page ingestion windows for historical reorg/move-regenerate pages, preventing empty page windows when updated_at still held creation time. Clarified current-index ranking and snapshot-based rollback. (#682)

2.1.2 - 2026-09-11

Changed

  • The managed routing snippet distinguishes a reviewed decision record kept in the repository (an ADR directory, a Keep the Why context/ tree) from a harness-local memory store: decisions go into the repo's record under its convention, ai-memory keeps recall, handoffs and session history and does not duplicate the record as a page. docs/usage.md ("Repo-native decision records") and docs/marker-file.md say to list such a directory in [capture] ignore_paths, and why (#700).
  • The managed routing snippet now states that Claude Code loads CLAUDE.md and does not read AGENTS.md: a project whose canonical instruction file is AGENTS.md needs a bare @AGENTS.md import line in CLAUDE.md, or the rules written there are absent from context at session start and reach the agent only if it opens the file. docs/install.md and docs/usage.md carry the same note beside the --target AGENTS.md guidance, and this repository's own CLAUDE.md now uses the import instead of a prose pointer (#680).

Fixed

  • A purge-session whose page-file cleanup failed was undone by the next watcher pass. The cleanup failure is reported in files_failed and leaves the database rows deleted while sessions/<id>.md is still on disk; nothing on the reindex path consulted the purged_sessions tombstone, so the reconcile tick 30 seconds later indexed the leftover file and the purged session's body was searchable again. The wiki reindex now skips a page whose session is tombstoned, the way it already skips a tombstoned scope, loading the tombstones once per directory per pass rather than once per page. The pass reports them as skipped_purged_sessions (#701).
  • The Docker wrapper (bin/ai-memory) now forwards GEMINI_API_KEY and GOOGLE_API_KEY into the container. Every other provider credential was on the -e forwarding allowlist, but these two were missing, so AI_MEMORY_LLM_PROVIDER=gemini (or the gemini embedder) reached the server while its key did not — the process then failed with provider not configured: GEMINI_API_KEY or GOOGLE_API_KEY even though the operator had exported it (#698).
  • serve no longer re-archives the whole data directory on every boot once the pre-migration backup receipt's archive has been deleted and auto-improve _pending/ sidecars exist. The OKF conformance scan that feeds the backup gate flagged those staging sidecars (which carry no frontmatter and are never migrated — SQLite owns their approval state) as nonconformant, so it kept falling through to a full archive. The scan now skips the project-root _pending/ subtree — a nested notes/_pending/ page still migrates — matching the watcher indexer and the existing ledger skip (#695, same class as #669).
  • A bare LLM_BASE_URL in the environment no longer redirects providers that talk to a fixed vendor endpoint. The variable is a cross-tool convention an operator exports once for a local Ollama, and ai-memory fed it to every provider: a gemini server then POSTed to http://localhost:11434/v1beta/models/<model>:generateContent and got Ollama's plain-text 404 page not found, surfacing as 502 Bad Gateway: {"error":"provider error 404: 404 page not found"} on bootstrap and consolidation. It now reaches only the dialects whose endpoint the operator supplies anyway — openai-compat, which has none without it, and opencode, whose Zen catalogue is an override — and is ignored elsewhere with a startup warn! naming the provider and the URL. An explicit llm_base_url (or AI_MEMORY_LLM_BASE_URL) still configures any provider, so proxying a vendor endpoint on purpose is unchanged. ai-memory llm-test resolves the base URL the same way, so it reproduces what serve will do instead of inheriting the same ambient override (#691).
  • Admin requests that fail for a reason the server owns are now logged server-side instead of existing only in the response body. POST /admin/bootstrap and the auto-improve routes serialized the error into JSON and told the log nothing, so an upstream provider failure could break every bootstrap while the log showed only the run starting — and a scheduled auto-improve tick, which has no client to print the body, failed with no operator-visible trace at all. A 5xx now emits a warn! naming the status, the operation, and the error. A 4xx stays quiet: the caller was told and the caller was at fault, so logging those would let any client fill the log at will (#692).
  • purge-session took the wiki mutation guard only for the file cleanup, after the database deletion had already committed. A watcher reindex could reinsert the still-present page in that gap and keep only the row, and a concurrent page write could lose its file to the cleanup. The purge now holds the guard across both steps, so reindexes, page writes and wiki moves finish before it starts and wait until it is done (#696, follow-up to #653).
  • Fixed empty native Codex tool observations by recognizing its top-level tool fields and preserving safe tool-family/call-ID metadata plus bounded, sanitized responses for recognized tools. Unknown tools and capture-excluded file operations retained their existing content restrictions; buffering, retry idempotency, and Stop/SessionEnd semantics were preserved. Corrected the install guide's outdated claim that Codex lacks native SessionEnd (#697).
  • Reads no longer report an empty project after the daemon restarts mid-session. The active-project pointer lives in process memory, so a restart — the one the packages' own post-upgrade note tells you to run — dropped it, and memory_status, memory_briefing, and every other unscoped read then resolved through the baked default scope and answered zero counts through the success path, with nothing in the log to distinguish "scope unresolved" from "project genuinely empty". serve now seeds a read-side fallback at startup from the most recently active project recorded in the database, bounded by the same TTL as a per-key entry, so a keyed miss right after a restart degrades to real data. The seed serves reads only: keyed per-actor entries are never reconstructed, an unscoped write resolves exactly where it did before, and the first hook event supersedes the seed — so pre-publish and eviction reads keep degrading exactly as they do today (#678).
  • The from-source AUR PKGBUILD now builds and tests on constrained AUR builders. Release LTO was disabled (options=('!debug' '!lto')) so the final link no longer gets OOM-killed on low-memory build hosts, and the check() step pins CARGO_HOME to the registry build() already populated before it repoints HOME at an empty test home, so the --frozen test run can resolve the packages it fetched instead of failing offline (#677).
  • CLI server errors now name the request. A non-2xx response from the configured server printed only server returned 404 Not Found: <body>, so a failure gave no clue which endpoint answered. The message now leads with the request method and path: GET /admin/open-sessions: server returned 404 Not Found: <body>. The error keeps the path only, so a token in the URL userinfo or query string never reaches a log line. (#679)
  • Wiki auto-commits stage what the wiki wrote instead of walking the whole tree, keep the repository open between commits, and no longer drop the commit when another session is writing a file at the same time. A session end now costs what it wrote, not the size of the wiki, and the git history no longer silently misses snapshots under concurrent sessions (#674).
  • The generated pi/omp and OpenClaw TypeScript integrations registered their session-end lifecycle event (session_shutdown for pi/omp, session_end for OpenClaw) with a synchronous, fire-and-forget handler. The host tears down the runtime as soon as a synchronous handler returns, killing the in-flight session-end POST before it completes, so the server never learned the session had ended: sessions.ended_at stayed NULL, and no summary or handoff was produced. The handlers are now async and await a bounded (2s) drain of the pending request(s) before returning, mirroring the joinable dispose-drain already used by the OpenCode integration (#676).
  • The bin/ai-memory container wrapper now matches multi-architecture image manifests against the host's platform architecture, eliminating a false-positive "a newer image is available on Docker Hub" warning on x86_64 and Podman. Additionally, emit_docker_run_script now preserves volume mount modes (such as :Z on SELinux/Podman environments) and filters transient runtime environment variables (HOSTNAME, container=podman). (#673)
  • ai-memory serve now stops on Ctrl-C and on SIGTERM, on both transports. The stdio transport listened for no signal at all, and the HTTP transport listened for SIGINT alone — so SIGTERM, what docker stop, docker compose down and systemctl stop send, reached no handler on either. What that cost depended on whether the server was PID 1. In the container it is (the image's ENTRYPOINT is exec form, with no init shim), and for PID 1 the kernel discards a signal whose handler is not installed: the signal was not merely unhandled, it was invisible, so docker stop sat out its whole grace period and ended in SIGKILL, docker kill was the only way out, and Ctrl-C on stdio did nothing at all. Everywhere else — under the native systemd unit, or a plain ai-memory serve in a terminal — the process is not PID 1, so the same signal fell through to the kernel's default disposition and killed it instantly instead, with no drain at all: the durable SessionEnd consolidation worker was cut off mid-flight rather than drained. Both transports now listen for SIGINT and SIGTERM, log which one arrived, and bound the drain at five seconds so a stateful or SSE MCP client holding a connection open cannot stall the exit — a stop that used to be instant and unclean now takes up to those five seconds and drains. The listeners are installed before the transport starts, so a signal arriving during a slow boot — migrations, the pre-migration archive — is handled rather than lost, and no container init shim (tini, docker run --init) is needed for the server to stop as PID 1 (#699).

2.1.1 - 2026-09-07

Changed

  • The managed routing snippet now states that ai-memory is the cross-harness memory of record: when the surrounding harness has its own local memory feature, durable project facts should be captured in ai-memory rather than duplicated in a harness-local store that other agents cannot see (#671).

Fixed

  • A nested .ai-memory.toml marker whose only content is a [capture] section (e.g. one that just sets ignore_paths) no longer resets scope resolution to default / basename. The native binary, the POSIX shell hooks, and the generated TypeScript integrations all used to resolve workspace/project and the other forwarded root-level settings (project_strategy, drop_subagent_captures, [recall] default_global, [briefing] keys) from the single nearest marker — the same one that decided [capture]/ignore_paths — so a capture-only marker in a subdirectory silently shadowed an ancestor marker's declared workspace/project and captures landed in the wrong scope. Scope and the other forwarded settings now resolve from the nearest marker that declares more than [capture], while [capture]/ignore_paths keeps reading the nearest marker unchanged — a marker that declares nothing else is scope/settings-transparent. Applies to the native ai-memory binary, the POSIX shell hooks (hooks/_lib.sh), and the generated TypeScript integrations for every adapter that resolves scope client-side (OpenCode, OpenCode2, pi, OMP, OpenClaw); outputs remain byte-identical otherwise. (#668)
  • Wiki auto-commits no longer re-hash the whole tree. Since the #594 guard, every commit cleared the git index and re-read every page, so a session end cost the size of the wiki and grew with it; the LongMemEval harness saw its ingest rate fall from 141 to 50 session ends a minute as the tree grew. Staging now goes through libgit2's stat cache and re-hashes the tree only when the tree write fails, which is the #594 case and is now covered by a test that removes a blob from the object store. Commits on one wiki are also serialized: two session ends at once used to collide on the index lock, and the losing snapshot was dropped with a warning. (#665)
  • Preserved typed relation edges in multi-page consolidation. Batch updates now include the same closed causes / fixes / contradicts schema as single-page consolidation and carry non-empty relations into wiki frontmatter and the link index, instead of silently discarding them. (#667)
  • OKF-conformed event ledgers are skipped by the indexer again, so a migrated store stops growing without bound. The reserved-file check treated any log.md / log-YYYY-MM.md carrying YAML frontmatter as an ordinary page, but the OKF v0.2 migration stamps frontmatter on every .md under wiki/, ledgers included. A migrated store therefore indexed its ledgers, and each hook append_event superseded them: one pages row per appended line, holding the whole multi-megabyte ledger body. One affected store reached 6,539 page versions and 14 GB from 101 live pages within four days of migrating. The check now reads past the frontmatter fence and classifies on the first body line, so a real page that happens to be named log.md is still indexed. (#660)
  • install-hooks --apply --capture-mode allowlist no longer captures repositories with no .ai-memory.toml marker on the five generated TypeScript integrations (pi, omp, opencode, opencode2, openclaw). The native ai-memory hook path already gated every lifecycle event on marker presence before spooling or sending anything (repository_admits_capture in ai-memory-hooks::capture_policy), but the generated adapters POST to <server>/hook directly and had only ported the ignore_paths denylist logic into their shared capturePolicy — never the allowlist admit gate — so an unmarked repository was still fully captured while the CLI printed that the protection was active. The shared template now bakes the selected --capture-mode into a CAPTURE_MODE constant and gates on marker presence (not on marker configuration state, so a marker with an empty [capture] section still opts a repository in) before any per-event disposition runs. The install-time "enforced"/"NOT in force" messaging is corrected to match: a script-fallback install is now the only path flagged as unenforced. (#661)
  • A UTF-8 BOM on a hand-edited wiki page that has no frontmatter no longer rides into the page body. markdown::parse stripped the mark before looking for the frontmatter fence, but the no-frontmatter path returned the untouched input as the body, so the mark stayed in front of the first line. Two things followed from that: derive_title no longer read the leading # as a heading, so the page was indexed under its filename instead of its title, and the OKF v0.2 file pass (which conforms frontmatter and leaves the body alone) re-emitted the mark after the closing --- fence, where it is a stray zero-width no-break space rather than a byte-order mark. Both paths now drop the leading BOM, which is what the frontmatter path already did. (#663)
  • serve no longer re-archives the whole data dir on every start once a monthly log ledger exists. The OKF conformance migration's nonconformant_files scan flagged every frontmatter-less log-YYYY-MM.md / log.md event ledger as a pre-OKF page (okf::is_conformant requires a type key and a ledger has none), so the pre-migration backup gate saw a non-empty pending list and took a full tar.gz snapshot on every boot — the flip side of #660, which taught only the watcher's indexer to skip ledgers by content. The migration scan now shares that same content-gated check (a reserved-looking filename is excluded only when its first body line is a hook log entry), moved into a shared ledger module so both call sites stay in sync; a page literally named log-2026-09.md whose body is prose is still migrated. (#669)
  • SessionEnd no longer writes an ephemeral sessions/<id>.md page for a session that logged no real work. is_lifecycle_only_session treated a session as skippable only when every observation was SessionStart or SessionEnd, so a single Stop observation — which OpenCode fires for purely internal work like branch-naming, alongside session.created + session.idle with no user prompt and no tool use — was enough to make the session look substantive and get a wiki page synthesized for it, flooding the wiki with no-op session pages. "Substantive" is now defined positively instead of negatively: a session counts as real work only if it contains a UserPrompt, PreToolUse, or PostToolUse observation (is_ephemeral_session in ai-memory-hooks). The atomic store-side check (end_lifecycle_only_session_in_tx in ai-memory-store) moved to the same positive kind IN ('user-prompt', 'pre-tool-use', 'post-tool-use') test so the two stay in agreement, and the PreCompact/PostCompaction checkpoint path gates on the same test rather than only observations.is_empty(). Provider-agnostic: this fixes the class for any harness that fires lifecycle-only Stop/Notification events, not just OpenCode. (#662)

2.1.0 - 2026-09-06

Added

  • Ordered LLM provider fallback chains (#648). [[llm_fallbacks]] in config.toml configures one or more additional providers, each with its own provider, model, optional base_url, and optional api_key_env (the environment-variable name holding that profile's key — never the key itself in config). The primary provider (llm_provider) always runs first; fallbacks run only after a transient failure (LlmError::is_transient(): 429, 5xx, timeout, connection error) advances past the current candidate, in declaration order, with the original request, JSON schema, and logical operation id preserved on every attempt. A deterministic failure (400/401/403/404/422, an unsupported schema, or a malformed response) still stops immediately — the same policy a single provider already had. Each candidate carries a 30s in-memory circuit: a transient failure opens it, a success closes it, and a restart clears all circuit state. Config::load validates every profile and resolves its credential once, at startup — a missing/empty provider or model, an unknown provider, or a RequiredApiKey provider with no resolved api_key_env fails startup rather than leaving a latent fallback that only fails once the primary is already down; a provider with a native credential source (OpenAI OAuth, Copilot) needs no api_key_env. GET /admin/status (and ai-memory status) now reports each candidate's provider/model label, last-selected/last success/error state, and circuit-open-until timestamp alongside the existing top-level LLM provider/model fields, which are unchanged for every existing single-provider deployment.
  • bootstrap --resume recovers an interrupted run from durable per-chunk progress instead of re-paying for every LLM call (#621). Each chunk's pages are recorded (keyed by a fingerprint of the pruned sources + chunk budget, so a re-run after new commits re-chunks and starts fresh) and --resume seeds them and skips those chunks; the wiki write stays atomic (nothing lands until the whole run completes) and progress is cleared on success. It adopts only the contiguous prefix of completed chunks and re-runs from the first gap or unreadable row (#635), so a resumed run can never diverge from a clean one by seeding a later chunk with context an earlier, still-missing chunk never produced. Complements the transient-retry from 2.0.3 (#617): retry reduces how often you need resume; resume covers the crash / restart case.
  • Added first-party OpenCode 2.0 beta (opencode2) support. install-mcp --client opencode2 merges the V2 mcp.servers remote entry (no enabled field, oauth: false for header credentials); install-hooks --agent opencode2 writes an ai-memory-opencode2.ts plugin in the beta { id, setup } shape (no runtime dependency: Plugin.define is identity), with handoff injection through the context hook and the same session.* lifecycle coverage as v1; ai-memory run opencode2 launches through the opencode2 binary and imports the beta session_v2/session_message transcript. The beta shares v1's config dir, session store, and agent kind, so no store migration ships. Verified live against beta-18999 (hook capture, handoff loop, managed launch/resume/import, opencode↔opencode2 session resume). Two beta caveats are documented: back up ~/.local/share/opencode first (the beta has migrated the shared database in place before), and move its background service off port 49374 when ai-memory serves there. Managed ledger-delta acknowledgement does not fire through the shared background service (redelivery, never loss); see the managed-workstreams notes (#622).
  • install-hooks --agent codex now wires Codex's SessionEnd hook, so a finished Codex session gets the same automatic end-of-session summary and cross-agent handoff Claude Code already gets. Codex shipped a first-class SessionEnd event in Codex CLI 0.145.0 (openai/codex#33895); the hooks/codex/session-end.{sh,ps1} scripts already existed and only needed adding to the Codex hook profile. On older Codex the event key is inert; ai-memory finalize-session --agent codex remains the fallback. merge_codex_payload no longer strips a SessionEnd key on every apply — a third-party SessionEnd hook is preserved alongside ours. ([#604])
  • AI_MEMORY_LLM_HEADERS (llm_headers in config.toml) attaches extra HTTP headers to every LLM chat request, for gateways that require a caller-identifying header. Entries are Name=Value or Name: Value, comma separated in the env var. Parsed and validated once at the configuration boundary, so a malformed entry fails at startup rather than on the first consolidation pass, and providers consume typed header material instead of operator strings. Headers ai-memory sets itself (authorization, content-type, x-api-key, x-goog-api-key, anthropic-version, anthropic-beta, openai-beta, host, content-length) are refused: reqwest appends rather than replaces, so a duplicate would break the request instead of overriding it. Values are marked sensitive on the wire and never logged — the Debug output carries header names only. ([#606])

Changed

  • Every LLM chat request now sends User-Agent: ai-memory/<version>. reqwest sends no user agent unless one is configured, so provider requests previously arrived anonymous, and gateways that require callers to identify themselves reported ai-memory as an unknown client. An AI_MEMORY_LLM_HEADERS entry for user-agent overrides it. The Copilot provider is unchanged: it keeps GitHubCopilotChat/<version>, the editor-plugin agent GitHub's Copilot API expects. ([#606])
  • x-opencode-session and the opencode user agent, both shipped in 2.0.2, are now operator-overridable like any other header: an AI_MEMORY_LLM_HEADERS entry for either name takes precedence over the provider's default. The default remains 2.0.2's — one LlmOperationId per logical operation, stable across retries and the strict/tolerant fallback — so nothing changes unless an operator asks for it. Override the session header to tell several ai-memory instances apart in OpenCode's metrics. ([#606])
  • The session-start brief's [briefing] max_chars now bounds the whole rendered brief, and its floor rose from 500 to 1500 chars. The scaffold the brief may never drop — the security notice plus both untrusted-history markers and the closing instruction — costs 832 chars by itself, so any budget under that could only ever have been honoured by dropping the boundary that marks the brief as untrusted. A project whose marker sets a smaller max_chars now gets 1500. (#657)
  • The build is self-contained on every platform: the vendored web stylesheet is now the default and TAILWIND_BUILD=1 cargo build -p ai-memory-web regenerates it, so no command needs TAILWIND_SKIP=1 any more. CI regenerates the bundle on Linux and fails if the committed static/tailwind.css is stale, a check that did not exist before.
  • Developer loop: cargo t (everyday, skips slow/stress modules) and cargo tf (everything) aliases over cargo-nextest, integration tests that compile into each crate's own test harness from tests/suite/ (78 test binaries down to 11 in the everyday loop; only the CLI keeps a separate one, and the evals harness builds only under --workspace), a dev profile that keeps only line tables, and an opt-in pre-push hook that runs the full tier. bin/release and the documented gate also run git diff --check. Measured: workspace edit-to-result ~380s to ~150s on macOS; the warm everyday test run 28s to 19s on a 32-thread Windows box.

Fixed

  • The bin/ai-memory container wrapper now runs on Podman without Docker (#636). It auto-selects the engine (AI_MEMORY_DOCKER override → docker → podman), defaults to the fully-qualified docker.io/akitaonrails/ai-memory image so Podman's non-interactive short-name resolution does not fail, and preserves the selected engine in the standalone-container recovery script ai-memory upgrade emits instead of hard-coding docker.
  • The OpenCode provider now sends gpt-5.6-luna requests to OpenCode Go's Responses endpoint. Luna is not served through Chat Completions, where plain and structured ai-memory calls returned HTTP 500 (#618).
  • AI_MEMORY_LLM_BASE_URL now works with AI_MEMORY_LLM_PROVIDER=opencode. The provider hardcoded OpenCode's Go endpoint and the factory dropped the configured base URL without a word, so Zen's general catalogue at https://opencode.ai/zen/v1 was unreachable through it — Zen and Go are separate products, not two spellings of one. Go remains the default, so existing setups are unaffected; an override keeps the x-opencode-session default and the user agent, since both endpoints correlate requests the same way. Model ids are per catalogue, so set AI_MEMORY_LLM_MODEL explicitly when overriding. OPENCODE_ZEN_BASE_URL is deprecated in favour of OPENCODE_GO_BASE_URL: the constant named Zen but has always held Go's URL. It keeps its value, so code compiled against it is unaffected. ([#606])
  • Fixed purge-session leaving the live sessions/<id>.md wiki file behind after deleting the session's SQLite rows. The admin endpoint now removes the scoped page file after the DB purge commits, reports actual cleanup through files_deleted / files_failed, and marks purge_session observer webhooks with partial_failure: true when file cleanup fails, so a later wiki reindex cannot resurrect the purged session. (#653)
  • [briefing] max_chars is now the size of the brief the agent actually receives. Page bodies were budgeted, but every section rendered after them escaped the accounting: the truncation notice, the crowded-out core page list, the recently-updated pointers, the closing untrusted-history fence and the agent footer. The escape pass that neutralises an untrusted <!-- ai-memory:untrusted-history --> marker inside a page body then lengthened the text by a further 6 chars per marker, after the count. A 4000-char budget rendered 6577 chars, and the 500-char floor rendered 3054 — six times what the operator asked for, on every opted-in session start. Bodies are now served first, the pointer sections are sized against what the bodies left (the omitted list degrading to a bare count when not even the paths fit), and the mandatory scaffold is reserved before a single body char is spent. The regression test asserting this was named render_session_brief_enforces_budget but only checked the brief's contents, never its length. (#657)
  • Authentication-disabled HTTP servers now ignore stale or unexpected Bearer headers and preserve anonymous access. Previously, a client retaining an old AI_MEMORY_AUTH_TOKEN received 401 Unauthorized even though the server reported auth=false; invalid Bearers remain rejected whenever static or human authentication is enabled. (#639)
  • Cursor sessions no longer land in the default default/scratch bucket. Cursor sends the workspace directory only as workspace_roots — its sessionStart / sessionEnd payloads carry no cwd key at all, and its tool events send cwd: "" — so cwd resolution produced nothing and the server fell back to its default project for every Cursor event. Both the native ai-memory hook path and the POSIX/PowerShell hook scripts now read workspace_roots (alongside Antigravity's workspacePaths) and treat an empty cwd as absent rather than as an answer.
  • Cursor sessions are no longer attributed to claude-code. The Cursor CLI also runs the hook commands declared in Claude Code's ~/.claude/settings.json, which install-hooks --agent claude-code hardcoded to --agent claude-code, so a Cursor-driven session was stored with agent_kind = claude-code. Hook payloads carrying Cursor's cursor_version marker are now attributed to cursor regardless of the ?agent= the hook command declared.
  • A project that first materializes while the server is running is now self-describing immediately, instead of only after the next startup backfill (#643). Scope manifests (_meta.md) were written at startup and on the rename/move admin paths, so a session in a checkout the server had not seen before produced a scope directory with pages but no manifest. Stop the server in that window and reindex could not rebuild that tree — the one situation where an operator most needs the rebuild to work. The manifest is now written with the scope's first page, one store lookup per scope per process, byte-identical to what the backfill writes so restarts still do not churn the wiki's git history. The startup backfill is unchanged and remains the repair path for trees written by older releases.
  • ai-memory reindex now names the exact missing or unreadable scope _meta.md path instead of collapsing the filesystem error to a bare No such file or directory (os error 2). This makes the existing startup-backfill workaround discoverable when a scope was first created during the server's last run. (#643)
  • Generated routing instructions and all project-scoped managed Agent Skills now distinguish session-aware MCP clients from static clients. Static clients are told to pass exact workspace + project values from .ai-memory.toml or operator configuration on every project-scoped call, preventing another session's last active project from capturing reads or writes; global searches and global preference writes retain their scope-free argument rules (#372).
  • Consolidation prompt assembly no longer re-renders the whole observation projection and re-scores every observation after each pruned one. With a few hundred long observations the quadratic loop cost ~14s per prompt; pruning now works from per-observation scores and block sizes computed once, with byte-identical output.

Security

  • A purged project or workspace can no longer be resurrected by a later reindex (#607, data-layer audit follow-up, item 2). purge_project / delete_workspace commit the DB deletion first and remove on-disk files afterward, best-effort; a crash or failure in that window left the markdown directory (_meta.md included) on disk with no row, and reindex rebuilt the scope from that manifest — silently undoing the purge. Each purge now writes a purged_scopes tombstone in the same transaction as the deletion (a whole-workspace delete uses a zeroblob(16) sentinel project id), and reindex_all skips any tombstoned scope instead of recreating it. The inert files are left for a later purge or manual cleanup; reindex stays non-destructive.
  • Concurrent writes to the same page path are now serialized (#607, item 3). Per-page writes take the shared side of the wiki mutation lock, so two writes to one (workspace, project, path) could interleave their file-rename and DB-upsert and transiently leave the on-disk markdown disagreeing with the DB is_latest row (self-healing on reindex, no data loss). A per-path async lock now serializes same-path writers while different paths still proceed concurrently; batches acquire their paths in a fixed global order so they cannot deadlock.

2.0.3 - 2026-09-04

Changed

  • Release, Docker, and the cargo install --git snippet now build with --locked (#628). Without it cargo install re-resolves dependencies and ignores the committed Cargo.lock, so a source install could silently pull a different schemars than the one tested — a plausible source of build-to-build structured-output schema drift. A unit guard also asserts the strict-schema normalizer keeps required equal to properties (and closes open maps), catching that drift class at build time rather than in production.
  • The _pending/auto-improve/ sidecar no longer renders a status: line (#624). It was written once at staging and could never change from pending — no code path updates or garbage-collects the sidecar — so it contradicted design principle 14 (SQLite owns approval status) and misled a human (or a second agent) reading the file. Status lives in ai-memory pending-writes list / SQLite; the sidecar is a staging-time snapshot.

Added

  • ai-memory status and the pre-migration backup log now report the data directory's filesystem free space (#629). The OKF-migration backup gate proves the archive is writable but said nothing about disk headroom — an operator whose 447 MB safety archive left 77 MB free had the store fail to extend its WAL eight minutes later, unlogged, for hours. This is pure signal (no threshold, no refusal): status shows filesystem free: beside the storage figures, and the "pre-migration backup verified" line now carries dest_free_bytes, so the same number is visible right before a migration.

Fixed

  • The pre-migration safety archive is now taken before the SQLite schema is migrated, so it is a genuine pre-2.0 recovery point a 1.x binary can reopen — restoring the documented "reversible upgrade" (#633). Previously the archive was written inside the wiki migration, which runs after Store::open has already advanced the DB schema, so the archived db/ was already at 2.x and a 1.x binary refused it — the documented rollback was impossible. The snapshot now runs in the boot path before Store::open, gated to the real 1.x→2.0 upgrade (a fresh install or already-migrated store takes nothing, unchanged) and idempotent; the OKF migration reuses that archive instead of taking a second one. An integration test asserts the archived DB reflects the pre-migration state.
  • Typed edges (relations) now actually work on OpenAI-family providers (#630). ConsolidatedPage.relations was an open map (BTreeMap<String, Vec<String>>), which the strict structured-output normalizer closes to additionalProperties: false — OpenAI strict mode cannot express arbitrary-key maps — so the model was structurally unable to emit any causes/fixes/contradicts edge on openai, openai-oauth, copilot, opencode, or openai-compat; the 2.0 typed-edges feature was silently dead there. relations is now a fixed three-field object (causes/fixes/contradicts), which strict mode expresses and which serializes to the identical relations: frontmatter — no migration, no downstream change. A schema guard pins the fixed shape.
  • finalize-session --agent hermes (and claude-desktop, crush, other) is accepted again (#623). These agents are captured and stored, but finalize-session reused the install-oriented agent enum — which is deliberately limited to agents with a first-party installer — so it rejected them at parse time and their sessions could never be closed, summarised, or consolidated. finalize-session is agent-agnostic, so --agent now accepts any agent the store recognises (a genuine typo is still rejected); install-hooks/setup-agent stay limited. A drift guard keeps the two sets from diverging again.
  • generated TypeScript integrations (--agent open-code, omp, pi, openclaw) authenticate again under install-hooks --apply. Since #552 the bearer is deliberately omitted from the rendered file and persisted to the 0600 <data_dir>/auth-token file, but only the native hook runtimes learned to read it back: the generated TS adapters kept rendering TOKEN = null with no runtime resolution, so every request they make (hook capture, spool drain, handoff fetch) goes out unauthenticated and is rejected by a Bearer-enabled server. The templates now render a resolveToken() helper (static embed, then AI_MEMORY_AUTH_TOKEN, then the auth-token file) used by authHeaders() and the spool writer. fetchHandoff() in the same adapters also ignored response.ok, so a 401 error body could be injected into model context as if it were a handoff; non-OK responses now return undefined like other failures.
  • bootstrap now retries a chunk's LLM call on a transient error before giving up, instead of letting one blip discard the whole multi-chunk run (#617). A provider 5xx/429 or a transport timeout/connect failure on chunk n previously aborted the run and threw away chunks 1..n-1 — the reporter's own runs died repeatedly to provider 520s and connection resets. Each chunk now gets up to 3 short, fixed-delay attempts; deterministic failures (auth, schema, a 4xx, malformed JSON) are not retried. The retry is deliberately short and bounded, not tenacity-style escalating backoff (cognee #2840). A durable --resume for the rarer crash/hard-failure case remains tracked separately in #617.
  • bootstrap no longer aborts the whole multi-chunk run when one chunk returns no pages key (#614). Later chunks are told which paths earlier ones wrote, so a model that judges the material already covered legitimately answers with a rationale and no pages; because Anthropic's tool_use schema does not enforce required fields the way OpenAI's strict: true does, that answer reached serde and failed deserialisation with missing field 'pages'. Since pages are only written after every chunk completes, the error discarded the work of all preceding chunks. pages now defaults to empty.
  • install-hooks --agent antigravity-cli on native Windows now emits an unquoted hook command (#611). The native-binary rendering wrapped the executable, --data-dir, and --server-url in double quotes, but Antigravity runs the string through cmd /c "<string>" without stripping inner quotes, so "C:\…\ai-memory.exe" was "not recognized", every hook failed, and the cp850 (non-UTF-8) error text aborted the whole agy session. Antigravity's command is now rendered bare; other Windows agents keep their quotes.
  • The wiki watcher no longer logs a scope-resolution failure for every page in a project directory the store has no row for, on every pass, forever (#613). The OKF v0.2 migration seeded an index.md into orphan directories (dirs with a valid-UUID name but no projects row — near-empty shells left by older history), and the watcher then warned once per file per 30s reconciliation pass indefinitely — hundreds of identical lines that buried real warnings, including, on one host, a 22-hour sqlite: disk I/O error. Reconcile now checks the project scope once per directory and skips an unresolvable one wholesale at debug, rather than retrying its pages; if the row later appears the directory indexes normally on the next pass. The same orphan skip now also guards the real-time directory-event path (reindex_project_dir), not just the periodic pass (#616).
  • The store error for an unknown project id now says "does not exist" instead of "does not belong to workspace X" (#612). A dangling id names no workspace mismatch to hunt for; the disambiguating lookup runs only on the failure path, so the common case still pays a single query.

2.0.2 - 2026-09-03

Fixed

  • Web UI: OKF bundle-index links (and any relative in-wiki link) no longer 404 (#603). Relative markdown link targets are now rewritten to project-scoped URLs — the same convention [[wikilinks]] use — instead of resolving against the web mount's <base href>, and a namespace (directory) path such as .../p/_lint/ now lists the pages under it rather than returning 404. External links, anchors, and dangerous schemes are untouched by the rewrite.
  • Pages with an empty frontmatter title no longer read back titleless or trip a bogus duplicate-title lint (#599). Auto-improve could store title: "" while the page had a proper # H1; now memory_read_page derives the title from the H1 (then the path stem) when the stored one is blank, the duplicate-title lint ignores blank titles instead of collapsing them into one Multiple pages share title "" finding, and the consolidator derives a title from the body H1 at write time so new pages never store an empty one. The DB title column (search) was already correct; this aligns the frontmatter-title readers with it without rewriting stored pages.
  • Local-embedding model download (fetch_model) now bounds a stalled connection instead of hanging startup forever (#602): a 30s connect timeout and a generous 600s overall ceiling (large model files, so not the inference clients' flat 120s that would fail a slow download).
  • The OKF v0.2 migration no longer crash-loops on startup with a libgit2 invalid object specified … class=Tree error that left the server unable to boot (#594). The wiki commit re-hashed files through libgit2's index, whose stat cache could trust a cached blob OID absent from the object database (a store carried across libgit2/git versions, or an interrupted earlier operation) — write_tree then aborted and, because the migration commits through that path, every restart failed. The commit now clears the index before staging so every file is re-hashed from the working tree. (The Docker runtime ships no git, so the fix is in-library, not a CLI fallback.)
  • HTTP headers that carry a credential in an opaque value are now redacted before reaching durable storage. Previously such a header matched no built-in pattern unless it used the Bearer keyword or an UPPER_SNAKE_TOKEN= shape. The bearer rule requires the literal keyword, and the generic env-var rule requires [A-Z][A-Z0-9_]*_TOKEN, which never matches a kebab-case header name. X-Amz-Security-Token (AWS SigV4), X-Api-Key, Private-Token (GitLab), and Ocp-Apim-Subscription-Key (Azure) all fall in that gap, and tool output echoing a curl invocation is a common way they reach capture. A key or token suffix on its own does not imply a secret, so it must be qualified by an auth word: Idempotency-Key, Continuation-Token and storage partition keys are left intact and stay readable in captured output.
  • install-hooks --agent zcode --apply no longer reports a hooks block ZCode has thrown away as no-op … (already up to date) (#600). ZCode validates hooks.events strictly and rejects the whole block, including ai-memory's own entries, over a single key it does not recognize, so capture never ran (hookCount: 0). Apply only ever merged the six event keys it writes, so a key it does not write was neither inspected nor reported, the file kept matching byte for byte, and the report claimed the install was current. Apply now also withdraws ai-memory's own entries from event keys it no longer writes, wherever they sit, and reports what it found: a warning naming any key it withdrew from, and a note for a key left unable to run anything, both naming the config file. Changed output: an unchanged ZCode config with such a key now reports no-op … (unchanged; see the notes below) instead of already up to date, so a redirected stdout no longer carries the reassurance without the cause. Event keys are never deleted and hooks ai-memory did not write are never removed, so a key holding someone else's hook is reported, not touched.
  • OpenCode Go requests now send User-Agent: ai-memory/<version> and a per-operation x-opencode-session header. Session consolidation and auto-improvement reuse the captured ai-memory session ID, while retries and structured-output fallback keep the same ID instead of appearing as unrelated requests (#608).
  • as_of time-travel queries and the entity retrieval stream now work on real stores. Both read the entity index, which was populated only from an LLM consolidator's entities: frontmatter — absent on the bulk of a mature store (bootstrapped pages predate it; stable pages are never re-consolidated), so the index sat empty and as_of returned nothing. Entities are now also derived from each page's tags: (the nouns a page is already labelled with) at write time, and a one-shot idempotent backfill runs on the first start after upgrading to populate the index for existing pages from their frontmatter — opening each entity-link window at the page version's own created_at so historical as_of is correct. Broad tags don't distort ranking (the entity stream is inverse-frequency weighted), and a store already written through the current path is a no-op. See docs/temporal.md.
  • The OKF v0.2 pre-migration backup walk now skips the .serve.lock single-instance lock and its .serve.lock.holder sidecar. On Windows the exclusive LockFileEx is mandatory, so the same serve process that runs the migration also holds .serve.lock and could not read it back while creating the backup — the walk aborted with os error 33 and the server crash-looped on every 2.0.x upgrade from a 1.x store (#593). Linux/macOS were immune because POSIX locks are advisory.
  • The Windows PowerShell wrapper (ai-memory.ps1) no longer crashes with git.exe : fatal: not a git repository … NativeCommandError when run outside a git repository — which broke ai-memory status and every non-repo invocation (#591). The wrapper's script-global $ErrorActionPreference = 'Stop' turned the repo-root probe's redirected native stderr into a terminating error under Windows PowerShell 5.1; the probe now runs under a localized SilentlyContinue, gates on $LASTEXITCODE, and restores the preference, falling back to the working directory as before. The unredirected Docker invocations were never affected.

Security

  • Handoff and workstream-event content is now size-bounded at the store boundary as defense in depth (data-layer audit follow-up, #607). The MCP/hook callers already scrub and cap this prose, but the store is the last gate before durable persistence, so it now bounds each handoff field (16 KiB, matching the observation body) and list length, and a workstream event's free-text content — a caller that ever forgot its own cap can no longer write unbounded content to the DB.
  • Untrusted relations: frontmatter values (relation keys and targets) are now length-bounded before being echoed into warning logs. On a shared server one caller's page is parsed by a process whose logs others read; a crafted or oversized relation key/target could bloat or pollute the log. Bounded to 200 bytes (UTF-8-safe), matching the bounding every other untrusted-content sink already uses. Found by a data-layer security audit of the store/wiki crates.

2.0.1 - 2026-09-02

Fixed

  • The human-readable wiki no longer drowns knowledge in machinery (UX audit after the 2.0 deploy). Three changes: the lint pass now supersedes a single _lint/report.md per project instead of minting a dated page every day (a long-lived store had accumulated 2,000+ lint pages — indexed, searched, and embedded; each pass now also prunes the legacy dated pile, and a clean pass removes the report entirely, so existing stores self-heal without a migration); the web project view splits machinery (lint reports, session captures, monthly logs, bundle indexes, _meta) into a collapsed System sidebar section and keeps Recent Activity to knowledge pages only — _rules stay with knowledge; and the homepage's LLM-optimised-memory explainer is now dismissible (persisted per browser) while the redundant always-on backup banner is gone — the migration dialog and ai-memory status carry that information.

2.0.0 - 2026-09-02

Added

  • Added a per-user launchd agent for macOS at packaging/launchd/com.github.akitaonrails.ai-memory.plist, the counterpart of the systemd user unit, shipped in both macOS release tarballs and documented in docs/macos.md. Until now macOS had no documented way to keep the server running once the terminal that started it closed, which silently stopped hook delivery. (#569)

    launchd expands nothing, so the template carries explicit __AI_MEMORY_BIN__ / __HOME__ placeholders rather than a home specifier, and passes neither --data-dir nor --config — both already resolve to the macOS platform defaults. ProcessType is pinned to Interactive because leaving it unset makes launchd throttle the job's CPU and I/O bandwidth, which the hook ingress budget cannot absorb. A bearer token, if one is configured at all, goes in the rendered plist's EnvironmentVariables: launchd has no EnvironmentFile equivalent.

  • status tells more of the truth (2.0 status audit): the wiki-format line (OKF migration state and the pre-migration backup archive while it exists), stored embedding triples by provider/model/dim, typed-edge counts by relation, and a write-queue depth gauge that surfaces a backpressured (wedged) writer — shown only when non-zero.

  • An opt-in cross-session "experience" pass ([auto_improve.scheduler] experience_every_sessions): every N newly completed sessions, the last few session summaries are reviewed side by side and knowledge visible only ACROSS trajectories — repeated workflows, re-stated preferences, architecture facts every session re-discovers, contradictions with stored decisions — is proposed through the identical schema-constrained, validated, eval-gated, pending-writes staging path as per-session auto-improve. Evidence must span at least two sessions; off by default and LLM-hosted, so the zero-LLM path is untouched. See docs/experience.md.

  • AI_MEMORY_EMBEDDING_PROVIDER=local — in-process sentence embeddings (pure-Rust candle BERT, all-MiniLM-L6-v2, 384-dim) with no API key and no external server. The ~87 MB model is fetched once into <data_dir>/models/ with source-pinned sha256s (offline installs drop the files in manually); vectors coexist with any provider's via the stored (provider, model, dim) triple, and switching is a config change. Feature-gated (local-embeddings, default on). The eval harness gained --embeddings local and publishes the hybrid LongMemEval row next to the zero-LLM baseline. See docs/local-embeddings.md.

  • The entity index carries ingestion-time validity windows (bi-temporal-lite): entity links open at their page version's creation and close when the version is superseded, backfilled for existing stores by an additive migration. memory_query gains an as_of ISO-8601 argument that turns the call into an entity-timeline lookup — "what did we know about X then" — returning the page versions valid at that instant, including ones superseded since. Ingestion time only, documented honestly as such. See docs/temporal.md.

  • Pages can declare typed relation edges in relations: frontmatter — a closed causes / fixes / contradicts vocabulary riding the existing links.link_type column (no schema migration). Declared contradicts edges surface as zero-LLM contradiction lint findings (including stale declarations whose target no longer resolves), and the session-end consolidation prompt may emit relations under the same schema-constrained vocabulary. See docs/typed-edges.md.

  • ai-memory export-okf --project <p> -o <bundle.tar.gz> and POST /admin/export-okf — stream one project's wiki as a validated OKF v0.2 bundle with a freshly generated index; a non-conformant page fails the export. Importing needs no command: unpack a bundle's concept files into a project's wiki directory and the watcher (or reindex) ingests them.

  • Added a LongMemEval retrieval benchmark to the evaluation harness (cargo run -p ai-memory-eval -- retrieval). It drives a real ai-memory serve subprocess end to end: haystack chat histories replay through POST /hook/batch at the production hook cadence, questions run through MCP memory_query, and results are scored session-level (hit@k / recall@k per question category, abstention questions reported separately). The 278 MB dataset is fetched on demand with a pinned sha256; baselines are published under docs/benchmarks/. The existing LLM A/B harness moved to the ab subcommand.

  • llm_reasoning_effort / AI_MEMORY_LLM_REASONING_EFFORT is a typed enum (none, minimal, low, medium, high, xhigh, max, ultra, persistent); unknown values fail at config load. Unset keeps the model default. Each chat provider maps the value to its native field: OpenAI reasoning_effort, OpenRouter reasoning.effort, xAI Grok reasoning_effort, Anthropic output_config.effort (plus adaptive/disabled thinking on models that accept it), and Codex Responses reasoning.effort. Gemini and Copilot ignore the key. Host-unsupported values are clamped (ultra/persistent → max on OpenAI/OpenRouter/Codex; Anthropic xhigh/max per model; none does not send thinking: disabled on always-on Claude models).

Changed

  • Hybrid retrieval is on by default: an install with no embedding_provider configured now gets in-process local embeddings best-effort — the model downloads in the background on first start (hybrid search enables on the next restart), existing pages are backfilled by a one-shot startup pass, and hosts that cannot fetch the model keep the previous FTS-only behaviour with a warning. Opt out with embedding_provider = "none"; configured providers are never overridden. Measured on LongMemEval-S: overall hit@5 rises from 0.617 (FTS-only) to 0.779 with local embeddings.
  • An unscoped memory_write_page / memory_handoff_begin from a caller whose active-project pointer does not resolve is now refused instead of written to the server's default project. The page such a write produced was real, attributed and searchable, and in a project nobody goes looking in. The error names both remedies — pass explicit workspace/project, or install the lifecycle hooks so the pointer is populated. Reads are unchanged: a read answering from the default is a wrong answer the caller can see, a write is a misfile they cannot. Callers with no identity coordinate at all (anonymous/legacy) keep resolving through the shared slot, and an install with no pointer information anywhere still resolves to the configured default ([#564]).
  • The wiki's on-disk format is now natively the Open Knowledge Format (OKF) v0.2: every page write fills the spec's type, generated, sources and stale_after keys (ai-memory's own fields ride along as spec-safe extensions), and each project directory is a conformant OKF bundle with an okf_version index. Existing stores are migrated automatically on the first 2.0 start — backup-gated: the entire data dir is archived to the user's home (or AI_MEMORY_BACKUP_DIR) and verified first, or the migration refuses to run; pages are rewritten in place (same ids, same version rows, timestamps untouched), and the wiki homepage shows where the archive is until it is deleted. A data dir migrated by a newer binary is refused by older 2.0+ binaries instead of silently mixing formats. See docs/okf.md and docs/MIGRATION-2.0.md.

Fixed

  • The generated TypeScript integrations (OpenCode, OMP, Pi, OpenClaw) no longer silently drop lifecycle events when the server is unreachable (#580). A failed or 5xx delivery is spooled to <data_dir>/hook-spool/ in the CLI's exact on-disk format — so ai-memory hook-drain and the shell hooks' piggyback drain deliver the backlog too — and each integration drains its own spool once the server is reachable again. Entries carry an idempotency key minted at spool time, so concurrent drains cannot double-ingest an event.
  • memory_read_page no longer ships a root-level anyOf in its tool schema (#577): the Anthropic Messages API rejects root combinators outright, so any provider-agnostic client routing tools through it (OpenCode with an Anthropic key, and every other Messages-API consumer) had its entire session 400 before a single tool ran. The exactly-one-of-path/query contract stays in the field descriptions and runtime validation; a new fence test asserts NO registered tool carries a root combinator so the class cannot return.
  • Docker-wrapper installs stage hook scripts back to the host-reachable ~/.local/share/ai-memory/hooks instead of the container's /data volume (#581): the #554 change made staging follow the resolved data dir, which inside the wrapper is a volume the host cannot execute from — settings pointed at container-only paths and every hook failed silently. Container runs now redirect staging to the wrapper's bind-mounted home contract; native installs keep following --data-dir/AI_MEMORY_DATA_DIR exactly as before.
  • Merging a project no longer reports a spurious content conflict when the only difference is generated.at: the merge comparison now uses the same modulo-timestamp projection as the store's idempotency rule. The old raw comparison made conflict detection timing-dependent — the identical page re-written across a second boundary 409'd as "different content" (the copy_purge_rerun_is_idempotent flake, root-caused via its diagnosable failure body on the Windows matrix run and now pinned by a deterministic cross-boundary test).
  • install-hooks now writes the capture-mode opt-in atomically. It was written in place, and every reader maps an unrecognised value onto denylist — the less private mode — so a crash, a full disk or a killed ai-memory upgrade mid-write left a truncated file that silently reverted an --capture-mode allowlist opt-in to capture-by-default, which is exactly what persist_capture_mode documents it must never do. The fallback itself is correct and stays; the torn write that triggered it is gone. Now routed through the same write_atomic (tmp + fsync + rename) the rest of the CLI already uses, per the project's atomic-writes invariant.
  • The 2.0 pre-release audit findings (post-merge adversarial review of the whole range): export-okf no longer fails on real deployments — scope manifests are OKF-typed at their writer (ending a startup tug-of-war that reverted the migration's typing on the same boot and self-healing manifests reverted by pre-fix binaries), approved auto-improve pages land conformant like every other write (also preventing phantom supersedes on the first reindex after a binary upgrade), and _pending/ proposal sidecars are excluded from the export walk. Pages retired WITHOUT a successor (decay tombstones, purge-regenerate, workspace merges) now close their entity-link validity windows so as_of cannot resurrect retired knowledge, with a V58 backfill for previously retired rows. The experience pass no longer burns its cadence window or stages an empty run when a scope has too few session pages. Backup destination guards canonicalize paths; as_of parse errors return invalid-params; the FTS5 probe connection is cached per thread.
  • status no longer overstates missing embeddings: empty-body pages — which no embedder can ever cover and the backfill skips by rule — are excluded from the "latest pages missing" figure and reported on their own line (observed live as a permanently stuck 427).
  • Natural-language searches no longer surface stopword trash: bare queries drop English stopwords before the FTS5 OR-join, so a page containing five "the"s cannot outrank the page whose content matches, and pages matching only stopwords no longer appear at all. Quoted phrases and explicit-operator queries are untouched, and an all-stopword query still searches its literal terms. Guarded by a CI-runnable search-quality suite asserting ranking usefulness, not just machinery.
  • memory_query no longer fails with fts5: syntax error when the query is natural language containing parentheses, apostrophe-quoted phrases, or a stray OR/AND (e.g. "my visit to the Museum of Modern Art (MoMA) and the 'Ancient Civilizations' exhibit"). Prepared FTS5 queries are now validated against the engine's own parser and degrade to an always-valid quoted bag of words when the preserved operator form does not parse; deliberate well-formed operator queries are preserved as before. Found by the new LongMemEval retrieval harness on its first full run.
  • ai-memory serve now takes an exclusive lock on <data-dir>/.serve.lock before opening the store, so a second server pointed at the same data directory refuses at startup and names the holding process instead of silently contending the single-writer actor, the wiki git handle, and the active-project pointer. The OS releases the lock when the process exits, so a crashed server never leaves the operator locked out; --force starts unguarded for an operator who knows the previous server is gone, and a filesystem that cannot lock at all downgrades the guard to a warning instead of refusing to start. (#563)
  • Preserved third-party string-form lifecycle hooks whose commands merely contain ai-memory or ai_memory when install-hooks --apply refreshes ai-memory's own entries. Legacy ai-memory script and native hook commands remain recognized by their specific hook signatures.