A self-organizing Obsidian vault that gives AI coding agents persistent memory.
TypeScript ai-agentsclaude-codecodex-cligemini-cliknowledge-managementobsidianobsidian-vaultpersistent-memorysecond-braintemplate
Changelog
v10.0.0 — 2026-10-12
Added
- make AGENTS.md the single instruction file
Changed
- say which issues are open for contributions, and require imperative PR titles
Fixed
- honor camelCase Stop hook re-entry flags
- build the mod report preface from the mod's name
- find frontmatter by its delimiter lines, not every ---
v9.1.0 — 2026-10-06
Added
- make /om-vault-upgrade ShardMind-aware, keeping local improvements through an upgrade
- add a prose-width rule so vaults stop teaching sessions to hard wrap
- gate every anchored promoted marker on resolving, and retarget markers on a split
Fixed
- run the release job's tests as CI does, and stop two Stop tests depending on qmd
- require shardmind 0.2.1 so a failed adopt cannot delete a clone's .shardmind files
- run Codex and Gemini hooks on Windows by a relative script path
- deliver every live goal at North Star's full level, and report any cut in the meter
- degrade eager-layer sections by levels instead of collapsing them whole
- resolve the QMD store path the way qmd does, honouring INDEX_PATH
- keep the test suite from writing QMD stores and configs into the user's folders
- enforce every tool's schema required fields in the MCP dispatcher
- link an informed_by title holding # or | by basename so the link resolves
- drive the frontmatter write check from frontmatter_required in the manifest
- drop backticks from memory filenames so a wikilink to them resolves
- refuse a record_work call with no title instead of writing an undefined note
- flag mcp_never_expose entries that are paths or folders instead of ignoring them
- derive the auto-memory directory on Windows so tidy-fix migrates strays there
- add a trailing flush so edits inside the QMD refresh debounce window get indexed
- report tidy-fix moves as fixed only when they succeed
- prefer a runnable qmd install over the first one that merely exists
- fail charcount when the section is not found, and match ## headings
v9.0.1 — 2026-10-05
Fixed
- require shardmind 0.1.7 so an older engine cannot overwrite edits on update
v9.0.0 — 2026-10-03
Added
- show the Stop report as one line under the answer through the mod
- deliver session context as an instruction file through a Claude Code mod
- add om_mod flags to session-start and stop-checklist
Fixed
- find the vault root in hook commands after /cd
- keep hook tests off the real QMD sentinel
- fail waiting qmd calls when the client's stdin breaks, instead of crashing
v8.6.0 — 2026-10-02
Fixed
- the Stop report shows the user a summary and reaches the agent with the next prompt
v8.5.0 — 2026-10-02
Added
- a changed Stop report reaches the agent, which decides whether to act
Fixed
- hold every hook writer under the output cap, with one cut routine
- hold the eager layer under Claude Code's hook output cap
- a Stop block carries the report once, as its reason
- compare Stop findings without their order, cap open loops on raw age
- route vault hygiene through actionable lifecycle hooks
v8.4.0 — 2026-09-02
Fixed
- the qmd search path was paying for a duplicate vector search and inheriting the wrong embedder
- the markup refusal's recovery is a different shape, not the one that folded
v8.3.6 — 2026-08-23
Fixed
- a wikilink terminator cannot survive into a memory filename (#237)
- a client that answers no roots/list can declare itself with OM_CALLER (#240)
- the first search waits out qmd's model download, and health round-trips a query
- the release job survives a re-run after a partial failure
v8.3.5 — 2026-08-17
Fixed
- the markup refusal names the fields the server received
- explain names the memory that superseded the withheld one
v8.3.4 — 2026-08-17
Changed
- give a question-shaped issue a route through the tracker
- use Obsidian's documented block-comment form in the template
Fixed
- audit what was refused, and serve a capture the tool refused
- a memory whose title outruns the slug is linkable by its title
v8.3.3 — 2026-08-17
Added
- /om-correct sweeps a corrected fact, and preserves the record
Changed
- state when a decision earns a record
- a new guard counts as landed once it has demonstrated a red
- Add Ko-fi funding link to FUNDING.yml
Fixed
- a retired claim is served only when its correction is served too
- a double-quoted alias with an escaped quote never matched its link
- a comma in a superseding title severs the correction link
- manifest-check compared platform paths to posix globs
- a superseded memory reaches only where its correction reaches
- the markup refusal names the failure, not only the field
- a CRLF note lost every alias, and three shapes lost some
- a YAML-escaped apostrophe in an alias never matched its link
v8.3.2 — 2026-08-03
Fixed
- emit the session-end checklist as JSON on every agent
- changelog generator no longer accumulates blank lines each release
v8.3.1 — 2026-07-31
Fixed
- health reports promotions that name a note but serve nothing
- the vault zip shipped a file left behind by a test
v8.3.0 — 2026-07-31
Added
- recall serves the promoted block, found by segmenting the note
- recall says when a memory has a corrected version it cannot serve
Changed
- name the predicate that answers "may this path be read"
Fixed
- a private marker past the head window read as public
- quote frontmatter as a YAML scalar, not by swapping quote characters
- never-expose withheld only the spelling you typed
- a capture could not be linked by the name it calls itself
v8.2.0 — 2026-07-30
Added
- warn when a memory claims general reach without saying why
Changed
- anchor vault consultation to an artifact, not to a moment
Fixed
- surface the cross-repo memory inbox in hygiene
- refuse a record_work capture whose arguments did not serialize
v8.1.1 — 2026-07-30
Changed
- ship the repo-side snippet, and make user scope the default install
Fixed
- resolve resource URIs that came from a search result
v8.1.0 — 2026-07-27
Added
- (reason) answer judgement questions by reading the vault in a second session
Changed
- (memory) cache the parse, and stop reading whole files to see their head
Fixed
- keep the engine's own tree out of the vault listing
- stop the meetings-inbox README flagging forever
v8.0.0 — 2026-07-26
Added
- reach the vault from any repo — MCP server and scoped cross-repo memory
Fixed
- derive the qmd index per vault instead of shipping a shared name
- point tsconfig at .scripts/ after the #149 move
- stop shipping tests for repo-only machinery
v7.0.1 — 2026-07-24
Added
- hold the SessionStart eager layer under a byte budget
Changed
- move scripts/ to .scripts/
- make env-export quoting unskippable
Fixed
- quote vault path in environment export
- release template leads with the ShardMind install path
v7.0.0 — 2026-07-19
Added
- tidy-fix — deterministic --fix consumer for detector findings
- phantom-edge findings ride the policyResults contract
- MEMORY.md as a rebuildable index (generator + regen tests)
- Write-Correctness Laws, agent-docs sync guard, index-first wrap-up, failing manifest-check
- classifier sub-hints — secondary discriminator on routing signals
- Recently Touched base + Stale Actives dashboard view
- eager-layer token efficiency — source-aware injection, spawn merge, hint dedupe, listing collapse
- wikilink parser lib + vault-wide broken-link ratchet gate
- om-tidy + hygiene-aware commands and agents (the acting half)
- vault hygiene — drift scan, write-time flags, config-driven open-loops
- memory-location guard, machine-readable policy results, phantom-edge warning
- QMD self-heal at SessionStart (empty store, ABI mismatch, version warn)
- injection-size meter for the SessionStart eager layer
- mermaid-visualizer and excalidraw-diagram skills
Changed
- dynamic command catalog, split doctrine, feature refresh across READMEs
- refresh vendored kepano/obsidian-skills to upstream a1dc48e
Fixed
- bump shardmind bootstrap fingerprint to qmd-v2 for the v7 migration
- derive qmd collection name from qmd_index (with legacy-name migration)
- skill review fixes — invalid mermaid fences, frontmatter wording, spacing
- normalize line endings and guard .gitignore against silent exclusions
v6.2.1 — 2026-06-02
Fixed
- tag the version-bump commit so released tags carry their own version
v6.2 — 2026-06-02
Changed
- migrate ShardMind hooks to split bootstrap/personalize lifecycle (#75)
- correct Codex skill discovery path (.agents/skills/ not .codex/agents/)
Changed
- migrate ShardMind hooks to split bootstrap/personalize lifecycle (#75)
- correct Codex skill discovery path (.agents/skills/ not .codex/agents/)
v6.1 — 2026-05-18
Changed
- surface macOS Obsidian CLI flicker caveat; refresh defuddle skill
Fixed
- extend isMarkdownFilename across remaining session-start call sites
- align qmd-bootstrap context-rm matcher with qmd's actual output
- stop qmd-bootstrap from mis-naming the collection on stock Windows
- stop SessionStart from flashing Obsidian app on macOS, close ExperimentalWarning stderr leak (#83)
v6.0 — 2026-04-26
Added
- install via ShardMind — wizard setup, modules, safe upgrades
Changed
- mkdir + cd before shardmind install
- use github: prefix in shardmind install / adopt commands
Fixed
- bump shard.yaml version on release
Added
- install via ShardMind — wizard setup, modules, safe upgrades
Changed
- use github: prefix in shardmind install / adopt commands
v5.1 — 2026-04-19
Added
- ARCHITECTURE.md, Obsidian-QMD ignore sync, lib cleanup, MCP-first
- keep QMD index fresh mid-session via detached hooks
- per-vault QMD named index, manifest-driven context, bootstrap
Fixed
- resolve qmd MCP failure on Windows via Node wrapper
- add .claude/scripts/.gitignore to block accidental node_modules commits
v5.0 — 2026-04-17
Added
- register QMD as MCP and teach Claude to consult brain topics via QMD
- inject brain topic index with empty-stub markers in SessionStart
Changed
- migrate hook and tooling layer from Python to TypeScript (v5)
- add roadmap section to translated READMEs (ja, ko, zh-CN)
- require issue-first for bigger contributions
- add roadmap section with Python to TypeScript migration notice
- Merge docs/upgrade-guide: upgrade guide, quick start, release improvements
- add upgrade guide and update quick start for multi-agent
Fixed
- close manifest coverage gaps and prepare v5.0 release metadata
- make hook commands cwd-independent via ${*_PROJECT_DIR:-.}
v4.0 — 2026-04-09
Changed
- Merge feat/multi-agent-support: add Codex CLI and Gemini CLI support
v3.8 — 2026-04-09
Fixed
- correct hook docs and add token efficiency section
v3.7 — 2026-04-08
Added
/om-meetingcommand — subject-forward meeting prep that searches the vault, surfaces open items and blockers, and brainstorms considerations beyond what's documented/om-intakecommand — processes raw meeting exports dropped inwork/meetings/, classifies content (1:1, project, decision, win, action item), and routes to the correct vault noteswork/meetings/inbox folder with README — staging area for raw meeting exports
Changed
- All 18 slash commands renamed to
om-prefix (e.g./standup→/om-standup,/dump→/om-dump) for discoverability — type/om-to see all commands via autocomplete - Updated all cross-references in commands, agents, scripts, hooks, and docs to use new
om-prefixed names - CLAUDE.md: updated command table (16→18 commands), added
work/meetings/to vault structure - README + all translations (ja, ko, zh-CN): updated all command names, agent tables, workflow examples, and added 3 new commands
brain/Skills.md: updated all command names, added/om-meetingand/om-intaketo Meeting Prep & Capture category.claude/settings.json: updated Stop hook checklist to reference/om-vault-audit.claude/scripts/classify-message.py: updated INCIDENT hint to reference/om-incident-capturevault-manifest.json: updated version fingerprints forom-prefix detection
v3.6 — 2026-04-07
Added
/prep-1on1command — prep for an upcoming 1:1 by loading person context, recent history, active work, wins to share, and a suggested agenda. Run before the meeting; use/capture-1on1after.
Changed
- CLAUDE.md: added
/prep-1on1to command table (15→16 commands) - README: added
/prep-1on1to commands table and vault structure diagram (15→16 commands) brain/Skills.md: renamed "Capture & Documentation" category to "Meeting Prep & Capture", added/prep-1on1
v3.5 — 2026-04-05
Added
- Demo GIF (
obsidian-mind-demo.gif) showing/standupand/dumpin action — embedded in all four READMEs
Changed
- README: restructured section order — Quick Start and Requirements moved up (from sections 12–13 to 4–5), Upgrading moved down (from section 11 to 16). No content added or removed.
- All translated READMEs (ja, zh-CN, ko) updated with same structure and demo GIF
v3.4.1 — 2026-04-05
Fixed
classify-message.py: hook output missing requiredhookEventNamefield — caused Claude Code to show "hook error" on every promptvalidate-write.py: same missinghookEventNamefield in PostToolUse output- Both scripts: added
OSErrorto stdin exception handling, wrappedmain()in top-leveltry/exceptfor resilience settings.json: bumped classify-message timeout 10s → 15s
v3.4 — 2026-04-05
Added
classify-message.py: native CJK classification patterns (Japanese, Korean, Simplified Chinese) for all 7 signal categories — pure CJK messages now trigger routing hintsclassify-message.py: English inflection coverage — all common verb forms (-ing,-s/-es) for delivery words now detected.claude/scripts/test_hooks.py: unittest suite forclassify-message.pyandvalidate-write.py— covers English, CJK, inflections, overlaps, false positives, type safety, integration
v3.3.1 — 2026-04-05
Fixed
classify-message.py: crash (rc=1) whenpromptis non-string type (int, None, bool, list) — now validates type before processingclassify-message.py:\bword boundaries failed with CJK text —\bdecision\bdid not match inのdecisionについてbecause Python treats CJK characters as\w. Replaced with Latin-letter lookaround(?<![a-zA-Z])/(?![a-zA-Z])classify-message.py: missing detection patterns — added1-1,praised,win,mentioned the,mentioned ato classifiersclassify-message.py: delivery words (shipped,launched,released,deployed,completed) now trigger both WIN and PROJECT UPDATE (categories were incorrectly mutually exclusive)classify-message.py: added new PROJECT UPDATE patterns —went live,rolled out,merged,cut the releaseclassify-message.py: changedlist[str]return annotation tolistfor Python 3.8 compatibilityvalidate-write.py: crash (rc=1) whentool_inputis null orfile_pathis non-string — now validates types before processingvalidate-write.py: translated READMEs (README.ja.md, README.zh-CN.md, etc.) were validated as vault notes — now skippedsettings.json: removedMultiEditfrom PostToolUse matcher (not a valid Claude Code tool name)settings.json: removed unsupportedmatcherfield from Stop hook (Stop event does not support matchers)settings.json: addedcompactto SessionStart matcher (hook was not firing after context compaction)settings.json: addedstop_hook_activere-entry guard to Stop hook (prevents infinite loop when Claude acts on checklist output)
Changed
classify-message.py: refactored from 7 hardcoded if-blocks to data-drivenSIGNALSlist — patterns are now declarative, overlaps between categories are explicit and visible
v3.3 — 2026-03-29
Added
/vault-upgradecommand — import and migrate content from an existing Obsidian vault (any version of obsidian-mind or arbitrary vaults). Detects source version via fingerprints, classifies notes, transforms frontmatter, fixes wikilinks, rebuilds indexes. Supports--dry-run.vault-migratorsubagent — classifies files (tiered heuristics: structural → frontmatter → content → fallback) and executes approved migration plans. Two modes: classification and execution.vault-manifest.json— declares template version, infrastructure vs user content boundaries, frontmatter schemas, version fingerprints, and field aliases. Enables version detection and targeted migrations.
Changed
- CLAUDE.md: added
/vault-upgradecommand andvault-migratoragent, updated counts (14→15 commands, 8→9 agents) - README: added "Upgrading" section explaining the migration workflow
brain/Skills.md: added/vault-upgradeto Maintenance category andvault-migratorto subagents table
v3.2.1 — 2026-03-29
Fixed
find-python.sh: detect Windows viauname -sand skippython3entirely (Windows Store stub is unreliable — can hang, consume stdin, or cause hook timeouts)find-python.sh: usecommand -vinstead ofpython3 --versionon macOS/Linux (faster, no side effects)classify-message.py: replace substring matching (in) with word-boundary regex (\b) — fixes false positives where "markdown", "wonder", "download", etc. triggered signalsclassify-message.py: addsys.stdout.flush()before exit to prevent buffered output loss on Windowsvalidate-write.py: remove unusedimport reand deadbodyvariable, addsys.stdout.flush()before exitpre-compact.sh: usefind-python.shinstead of hardcodedpython3(was bypassing the cross-platform resolver), merge two Python calls into onesession-start.sh: quote$CLAUDE_PROJECT_DIRin exported value (paths with spaces broke the export)session-start.sh: exclude.git/from vault file listing (consistent with other exclusions)charcount.sh: use${1:-},${2:-},${3:-}for positional args (withset -u, missing args crashed before reaching the friendly usage message)
v3.2 — 2026-03-29
Added
/humanizecommand — voice-calibrated editing that matches your writing style, not a generic AI word blacklist/weeklycommand — cross-session weekly synthesis with North Star alignment, pattern detection, and uncaptured win spotting
Fixed
validate-write.py: normalized path separators for Windows (backslashes weren't matching forward-slash skip list)validate-write.py: addedthinking/to skip list (scratchpad notes shouldn't trigger validation warnings)
Changed
- CLAUDE.md: reordered command table by category, added new commands, fixed stale counts (10→14 commands, 7→8 agents), added
review-fact-checkersubagent - README: updated command table, daily workflow section, command and agent counts
brain/Skills.md: added Editing & Synthesis category, new commands, usage notes, and Weekly Review workflow
v3.1 — 2026-03-27
Added
- Vault-first memory system — all project memories live in
brain/(git-tracked),MEMORY.mdbecomes an index-only pointer /self-reviewcommand — guided self-assessment workflow with strategic calibration, fact-checking, and character limit validation/review-peercommand — peer review writer with visibility classification, tone rules, and quality checksreview-fact-checkersubagent — verifies every claim in a review draft against vault sourcescharcount.shutility script — counts characters in markdown sections for review tools with character limits.claude/memory-template.md— template users copy to~/.claude/to wire up vault-first memory
Changed
- CLAUDE.md: "Two Memory Systems" replaced with "Memory System" (vault-first rule, setup instructions)
- CLAUDE.md: Rules section updated to enforce vault-first memory (never create files in
~/.claude/) - README: updated memory description, command/agent counts, added new commands and subagent
brain/Skills.md: added new commands, subagent, and updated review cycle workflow
v3 — 2026-03-26
Added
/standupcommand — morning kickoff that loads context and suggests priorities/dumpcommand — freeform capture that auto-classifies and routes to the right notes- 7 subagents:
brag-spotter,context-loader,cross-linker,people-profiler,review-prep,slack-archaeologist,vault-librarian - 5 lifecycle hooks: SessionStart (rich context injection), UserPromptSubmit (message classification), PostToolUse (write validation), PreCompact (transcript backup), Stop (session end checklist)
- QMD semantic search integration (optional) with custom skill in
.claude/skills/qmd/ - Hook scripts in
.claude/scripts/:session-start.sh,classify-message.py,validate-write.py,pre-compact.sh thinking/session-logs/for transcript backups before context compaction
Changed
- README rewritten as product documentation with badges, scenarios, daily workflow, and performance graph sections
- CLAUDE.md updated with subagents table, hooks table, QMD skill reference,
/standupshortcut in session workflow brain/Skills.mdreorganized by category (Daily, Capture, Performance, Maintenance) with subagents and hooks tables
v2 — 2026-03-26
Added
Home.md— vault dashboard with embedded Base viewsbases/— 7 centralized Obsidian Bases (Work Dashboard, Incidents, People Directory, 1-1 History, Review Evidence, Competency Map, Templates)work/active/+work/archive/YYYY/— explicit project lifecyclework/incidents/— structured incident trackingwork/1-1/— 1:1 meeting notesorg/— organizational knowledge (org/people/,org/teams/,People & Context.md)reference/— codebase knowledge and architecture docsperf/evidence/— PR deep scans for review prepperf/brag/— quarterly brag notes- 8 slash commands:
/peer-scan,/slack-scan,/capture-1on1,/vault-audit,/review-brief,/incident-capture,/project-archive,/wrap-up .claude/update-skills.shfor syncing obsidian-skills from upstream
Changed
- Renamed
claude/→brain/with split files (Memories index, Key Decisions, Patterns, Gotchas, Skills, North Star) - Moved
perf/Review Template.md→templates/Review Template.md - CLAUDE.md rewritten with comprehensive session workflow, note types, linking conventions, Bases documentation, properties for querying, agent guidelines
perf/Brag Doc.mdupdated to quarterly sub-note structure
Removed
claude/Memories.mdmonolith (replaced by split brain/ files)
v1 — 2026-03-01
Initial release. Basic vault structure with:
claude/— Memories, North Star, Skills (monolithic)work/— flat work notes with Index.mdperf/— Brag Doc, Review Template, competencies/templates/— Work Note, Decision Record, Thinking Note, Competency Notethinking/— scratchpad- SessionStart hook (file listing injection)
- obsidian-skills pre-installed