Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
44.8 KB
Raw Blame

Multi-source brains

A single gbrain database can hold multiple knowledge repos. Each one is a source: a logical brain-within-the-brain with its own slug namespace, its own sync state, and its own federation policy. The rest of this guide walks the three canonical scenarios.

(Sources are the within-one-database axis. If you want to connect a whole separate database — a team-published brain with its own access policy — that's the brain axis: gbrain mounts add. See docs/architecture/brains-and-sources.md for the two-axis topology.)

The three scenarios

1. Unified knowledge recall (wiki + gstack)

You have a personal wiki and a gstack checkout. Both belong to you, both are knowledge you want your agent to recall across. When you ask "what did I learn about X?" you want the best hit whether it lives in the wiki or in a gstack plan.

# Register the gstack source, federate so it joins cross-source search
gbrain sources add gstack --path ~/.gstack --federated

# Pin the directory so `gbrain sync` knows which source it's walking
cd ~/.gstack && gbrain sources attach gstack

# Initial sync
gbrain sync --source gstack

# Now `gbrain search "retry budgets"` returns hits from BOTH wiki and
# gstack. Each result includes source_id so the agent can cite properly.

Result: wiki pages and gstack plans are separate (different source_ids, different slug namespaces) but share the search surface.

2. Purpose-separated brains (news-desk + essays)

You run two completely different content pipelines on the same backend. news-desk covers company news and founder profiles. essays is personal writing. You explicitly DON'T want them mixed in search — news content leaking into essay searches is a bug, not a feature.

# Two sources, both isolated (federated=false)
gbrain sources add news-desk --path ~/news-desk --no-federated
gbrain sources add essays --path ~/writing --no-federated

# Pin each checkout directory
(cd ~/news-desk && gbrain sources attach news-desk)
(cd ~/writing && gbrain sources attach essays)

# Sync each independently
gbrain sync --source news-desk
gbrain sync --source essays

Result: searching from neither directory returns the default source (your main brain). Searching from inside ~/news-desk returns only news-desk hits. Searching from inside ~/writing returns only essays. Federation is opt-in, not leaked.

To search across every source explicitly on demand (trusted local CLI; --source takes exactly one id):

gbrain search "tech layoffs" --source-id __all__

3. Mixed (wiki federated + sessions isolated)

Your main wiki is federated with a few trusted sources. Your session transcripts (gbrain transcripts ingests them) land in a separate isolated source so they don't dominate every search result.

# Federated sources
gbrain sources add gstack --path ~/.gstack --federated

# Isolated source for session transcripts
gbrain sources add sessions --path ~/.claude/sessions --no-federated

Resolution priority

When any command needs to pick a source, gbrain walks this list (highest first):

  1. Explicit --source <id> flag.
  2. GBRAIN_SOURCE environment variable.
  3. .gbrain-source dotfile in CWD or any ancestor directory.
  4. A registered source whose local_path contains the CWD (longest prefix wins for nested checkouts).
  5. The brain-level default set via gbrain sources default <id>.
  6. The seeded default source.

So inside ~/.gstack/plans/ on a brain that pinned gstack to ~/.gstack via .gbrain-source, gbrain put implicitly writes to the gstack source. Outside any registered directory with no env/dotfile set, it writes to the default.

Federation flag

Every source row stores config.federated: boolean in its JSONB config.

Value Meaning
true Source participates in unqualified gbrain search "X" results.
false or unset Source only searched when explicitly named via --source <id> or qualified citation.

A new source added without --federated or --no-federated is federated on a brain only you use (access.serves is one, brain.audience reads personal, and you are not on a thin client). On a shared brain it stays unset, and sources add and sync --source <id> print a source_not_in_default_search notice whose fix is gbrain sources federate <id>, marked ask_user: it asks whether the source should be shared with every connection that reads default search, and a source meant to stay private (a member's personal source, say) needs nothing. Sources you added before this change keep their setting.

The seeded default source is federated=true so single-source brains behave as you'd expect — every page appears in search.

Flip later with gbrain sources federate <id> / unfederate <id>. Federating a source (or restoring it, which re-federates) lets every connection with no source grant read it, including pages a member-bound credential wrote there without readable_by. The command still federates, and it lists the pages a remote token or OAuth client wrote there without readers (count, a few slugs with the member each credential is bound to now, pages_exposed under --json) with gbrain access who-can-read to check one and gbrain access rule add to keep a folder to its member.

An unscoped connection whose sources hold no pages

A remote token minted without --sources, or an agent stdio connection with no binding, has no source grant: it reads the default source plus every federated source. Because a new source is not federated, a brain whose pages all live in such sources would answer that connection's every search with an empty result that reads like a clean miss.

So an unscoped search, query, recall or list_pages from such a connection that comes back empty is refused with sources_not_readable when the sources it reads hold no page that every reader of the source can read (for recall and query, no active world fact either) while other sources on the brain hold pages. Quarantined pages count on neither side, and neither do pages kept to named readers (a member's own pages included) or private pages, so a member can be refused while its own pages sit in default. The check is one yes/no about the whole brain, independent of the query: whether another source holds pages. It never says another source matches the question, and it never names a source or page the connection cannot read. Its fix is host-side: the user picks the source these connections should read and the brain host runs gbrain sources federate <source>, or grants the connection its sources (gbrain auth rescope --token <name> --sources <ids>).

Nothing else changes. A read whose sources hold pages but match nothing answers empty, as does a brain where no source holds a page. A token granted sources, an OAuth client, a bound connection and a read that names source_id or source_ids are never refused: the refusal only ever stops a call, never widens what it reads. Federating a source never shows its visibility: private pages to a remote connection.

On a brain that is not managed, gbrain sources add without --federated or --no-federated warns on stderr that the new source is not federated, and when the sources an unscoped connection reads hold no page it also prints the next step on stdout (next: gbrain sources federate <id>), and as next (command, argv, why) under --json. A managed brain's sources add prints its lifecycle receipt instead, without the warning or the next step.

Explicit reads from a bound agent connection

An agent connection started with GBRAIN_SOURCE=<id> (or inside a directory pinned by .gbrain-source) is bound to that source. Its unqualified reads stay on the bound source. It may still name another source in an explicit source_id read (search, query, get_page, list_pages, resolve_slugs, recall, get_links, get_backlinks, traverse_graph) when that source is federated=true; the result holds only the named source's rows. Writes still go to the bound source only.

Refusal hint Why Fix (on the brain host)
<id> is not federated The named source is not federated. gbrain sources federate <id>, or start the connection without the binding.
<id> opted out of federation The named source was unfederated. gbrain sources federate <id> if it should be readable from other sources.
<bound> opted out of federation …, so it reads no other source The bound source itself is isolated (federated=false), so it never reads another source. gbrain sources federate <bound>, or start the connection without the binding.
Your token is not granted <id> An HTTP token or OAuth client whose grant does not include the source. gbrain auth rescope-client <client_id> --federated-read <ids> for an OAuth client; gbrain auth rescope-token <name> --sources <ids> for a legacy bearer token.

search_by_image, open_loops and the code-intel tools keep their stricter rule: an explicit source_id must be inside the connection's own source or grant.

Narrow one request with source_ids

search, query, recall, memory_recall and list_pages take source_ids, a list of source ids (CLI: --source-ids source-a,source-b). The request reads only those sources:

  • Each id is checked exactly like an explicit source_id. A remote agent naming a source outside its grant gets permission_denied (scope_denied from recall) for the whole request, and a removed or archived source gets unknown_source (not_found from recall). So the list can narrow a request but never widen it.
  • Omitting source_ids, or passing [] or a blank string, reads every source the caller can read, as before.
  • With source_id as well, the request reads the sources in both (source_id: "__all__" keeps the list). If none are left, nothing is searched: the result is empty with an empty_retrieval notice, never a fallback to the wider scope.
  • __all__ and malformed ids in the list are invalid_params.
  • On the starter surface, recall and list_pages accept source_ids but leave it out of their listed schema to keep the tool list small; harness code sets it, and request_tools returns the full schema.

Every part of the request reads the narrowed sources: keyword, vector and title matches, other names, entity anchoring, saved facts, the CRAG re-run, think escalation and evidence delivery; on recall, every fact arm and the page-search arm.

Say to your agent: "Search my brain for the acme-example renewal, but only in source-a."

A bot acting for many members

A shared brain often holds one source per member or project, with one bot answering for several people. Give each member their own token, granted only the sources that member may read, and send each member's requests with that member's token. This is the supported production setup today: one serve-http API key per member, each with its own source grant. The server never learns who is asking from the request itself.

gbrain auth create bot-alice-example --member alice-example
gbrain auth rescope --token bot-alice-example --sources source-a
# OAuth client instead of a bearer token:
gbrain auth rescope --client <client_id> --sources source-a --member alice-example

Once the brain has two or more enabled members (or config access.serves is several), a remote read through a token, OAuth client or stdio server bound to no member is refused with requester_scope_required, and gbrain auth create refuses an unbound token. A connection that existed when the brain started serving several people reads for 14 more days, with a notice on every answer (_meta.access.requester_scope) naming the date and the command that binds it. Bind every member's token with --member. A brain for one person sees no change, however many sources it holds.

The token's grant is the security boundary: the server scopes every call to the grant of the token that made it, whatever the call's arguments say. source_ids is a convenience inside that boundary. A bot holding one token granted several sources can pass source_ids to keep one request to the current member's sources, but a prompt-injected or buggy bot can omit it, so it does not separate members who must not see each other's pages. Use per-member tokens for that, and source_ids to narrow further.

A trusted system that names the person on each call (QM's x-qm-actor) can share one client instead: grant it every member's source and bind each source to its member (gbrain access member source add alice-example dm-alice-example). An attested call then reads the grant minus sources bound to other members, and binding never adds a source to a grant (QM guide).

Lay out sources by who may read them

Source grants are the coarse boundary: without page readers, two pages in one source are readable by exactly the same tokens. Inside one source, a page that names readers in frontmatter readable_by is read only by credentials bound to a listed member, and a folder rule (gbrain access rule add) does the same for a whole prefix (Model C); both only narrow a grant, never widen it. Either way, put content in sources by who may read it, before any member's bot connects:

  • One source per channel, team or member. Give a channel's pages a source of their own (channel-a), a team's documents theirs (team-a), and each member a private one (private-alice-example). Content everyone may read goes in a shared source (public).

  • One grant per member. Grant each member's token or client every source that member may read, and nothing else:

    gbrain auth rescope --token bot-alice-example --sources private-alice-example,channel-a,team-a,public
    gbrain auth rescope --client <client_id> --sources private-alice-example,channel-a,team-a,public
    

    The first id is the token's write source, so a member's saved memories land in their own private source. Keep that source unfederated: federating it later makes the pages already written there readable by every credential without a source grant, since they carry no readable_by.

    A credential bound to a member (--member) whose write source is default (a token with no --sources or with --sources default, or an OAuth client left at its registration default) or a federated source writes where every credential without a source grant reads. A page such a credential creates is therefore written with readable_by: [member:<id>], readable only by that member (and the local CLI); rewriting it without readable_by keeps those readers, and the member shares it by naming readers (readable_by: [alice-example, group:team-a]). An edit to an existing page that names no readers leaves it shared. POST /ingest refuses such a credential. Pages derived from such a page (extracted atoms, chronicle events, derived_from pages, and a synthesized concept that uses it) are readable by no remote caller, the writer included. A page teammates should read (a shared people/ or companies/ page, or a page in a folder a rule scopes to a group) names its readers when it is created; otherwise the first member to create a slug keeps it, and every other remote credential writing that slug gets page_not_found until the operator adds readable_by on the brain host. Before this release those pages named no readers, so other members and unbound tokens read them: check pages a member wrote with gbrain access who-can-read <slug> and add readable_by on the brain host to any that should stay private.

  • Let each member read their own private pages (optional). Remote reads hide every visibility: private page by default, even in a granted source. To let a member's agent read the private pages in the sources its grant names, and no others, opt that token or client in:

    gbrain auth rescope --token bot-alice-example --private-pages granted-sources
    gbrain auth rescope --client <client_id> --private-pages granted-sources
    

    A token with no source grant, a federated source the grant does not name, and a request's source_ids never add a source. The grant applies to search, query, recall, entity and the page, chunk, version, timeline, link and graph reads; context_pack, delta, think, recorded-time reads (known_as_of) and writes keep hiding private pages even in a granted source, so a page the agent can open may read as missing there. In a named source the grant also opens pages with a privacy alias (access, confidentiality), atoms, concepts and events derived from private or unknown origins, and pages whose declared lineage is private; pinned questions stay owner-only. --private-pages none hides them again on the next request, and gbrain auth list shows each token's setting. A harness token rotation (gbrain setup, gbrain bootstrap harness) does not carry the grant; it prints the rescope that grants it again.

  • Let a member's agent read private facts (optional). Facts saved private stay hidden from remote reads by default. gbrain auth rescope --token bot-alice-example --private-facts granted-sources (or --client <client_id>) lets that one credential read them in the sources its grant names, and no others. It never opens a fact held on, provenanced from or about a private, readable_by, folder-ruled, quarantined or member page, a fact with no page to check, or a transcript fact; context_pack, delta and writes stay world-only. --private-facts none turns it off on the next request. Details: MCP administration.

  • On a personal brain, name the owner instead. gbrain access owner set <member> lets every credential bound to that member read the private pages in all the sources its grant reads. What each designation means for each caller: access designations.

Warning: one shared source has no boundary inside it. A bot that answers everyone with one unscoped token over a single source exposes everything in that source to every requester. Neither source_ids nor a search limit can narrow a read inside one source: source_ids chooses whole sources, and a limit only shortens a ranked list, which can still start with a page the requester should not see. Semantic search makes this more visible, because it returns related pages that a keyword search would not match. Split the content into sources first, then grant per member.

Links, backlinks and the graph across sources

gbrain links, gbrain backlinks and gbrain graph (MCP: get_links, get_backlinks, traverse_graph) read the same sources as search for the same caller:

Caller Unqualified read --source <id> / --source-id <id> --all-sources
Local CLI, source not pinned Every federated source: one read per source, merged That source Every source
Local CLI, pinned by --source, GBRAIN_SOURCE or .gbrain-source The pinned source That source Every source
Remote agent with no source grant Every federated source A federated source Every federated source
Remote agent with a grant Its granted sources (never wider) A granted source Its granted sources
gbrain backlinks companies/acme-example                     # resolved scope
gbrain backlinks companies/acme-example --source business   # one source
gbrain links people/alice-example --all-sources --json      # every source
gbrain graph people/alice-example --depth 2 --all-sources

Each call returns at most 500 rows (--limit N, up to 5000), merged across sources and identity members in link order before the cut. When more match, the last row carries truncated: true and next_cursor; continue with --cursor <next_cursor> and the same flags, or pass --all to page through every row.

get_links keeps working as an alias of links. Passing --source together with --source-id or --all-sources is an error. A source that does not exist or is archived answers unknown_source; a remote agent naming a source outside its grant or federated set gets permission_denied with the fix, even when that source is archived. Private pages stay hidden from remote agents exactly as in search.

Graph output carries source ids: each gbrain graph node has source_id, and each edge returned to agents has from_source_id / to_source_id. The same slug in two sources is two nodes or two edges.

When a local unqualified read finds nothing but the page has links in a source outside the read (a non-federated source), the CLI prints the per-source counts and the exact rerun command on stderr, for example gbrain backlinks companies/acme-example --source priv; stdout stays the empty result and --json prints no hint. Remote agents never get this hint, so a source they cannot read stays indistinguishable from an empty one.

On a thin-client install the link commands send an ambient GBRAIN_SOURCE or .gbrain-source binding as source_id, so they are narrowed the same way search is. An empty narrowed result says which binding scoped it and prints the --all-sources rerun command. If the brain host is older and ignores source_id or --all-sources, the command fails with "the brain host does not support … upgrade the host" instead of printing an unscoped result. Hosts that predate unknown-parameter warnings cannot be detected; upgrade the host first.

Say to your agent: "Show me everything that links to acme-example across my sources."

Say to your agent: "Which pages does alice-example link to in the business source only?"

Commands

The most-used subcommands (run gbrain sources --help for the full, always-current reference — it also covers status, current, set-cr-mode, and the push/pull durability surface).

Changing the contextual retrieval mode, per source with gbrain sources set-cr-mode <id> <mode> or brain-wide with gbrain config set search.contextual_retrieval <mode>, does not restamp existing pages: each keeps the mode recorded when it was imported, and the new mode applies to pages imported from then on. Both commands print this after the write, with gbrain repair contextual-mode [--source <id>] --apply for pages that have no recorded mode (preview it first without --apply; see repair).

gbrain sources add <id> --path <p> [--name <n>] [--federated|--no-federated] [--force] [--json]
                               Register a source. id: [a-z0-9](?:[a-z0-9-]{0,30}[a-z0-9])?
                               --path must be a git repo (or a subdirectory of one) — see
                               "The git requirement for --path sources" below. --force
                               registers a plain directory: sync imports its files without
                               Git, and a file removed from it keeps its page.
gbrain sources add <id> --url <git-url> [--pat-file <p>] [--clone-dir <path>] [--no-harden]
                               Clone + register a remote repo in one step; auto-hardens
                               for durability when a PAT is provided (see "Durability" below).
gbrain sources list [--json]   List all sources with page counts + federation state.
gbrain sources archive <id>    Soft-delete: hide from search, keep data for a TTL
                               grace window. Prefer this over `remove`.
gbrain sources restore <id>    Un-archive. `gbrain sources archived` lists expiries;
                               `gbrain sources purge` permanently deletes expired archives —
                               except sources still referenced by a registered OAuth client
                               (reported as `Blocked:`, sweep continues); revoke or rescope
                               the client (`gbrain auth revoke-client <id>`) and re-run.
gbrain sources remove <id> [--confirm-destructive] [--dry-run]
                               Permanently cascade-delete a source (pages, chunks,
                               timeline). Shows an impact preview first.
gbrain sources rename <id> <new-name>
                               Change display name only; id is immutable.
gbrain sources default <id>    Set the brain-level default.
gbrain sources attach <id>     Write .gbrain-source in CWD (like kubectl context).
gbrain sources detach          Remove .gbrain-source from CWD.
gbrain sources federate <id>
gbrain sources unfederate <id>
gbrain sources mirror-readonly <id>
gbrain sources mirror-writable <id>
gbrain sources record-origin <id>... --remote|--local [--apply] [--json]
                               Record who added older URL sources (doctor
                               url_source_origin); preview unless --apply. See
                               docs/guides/repair.md#url-source-origin.
gbrain sources refresh <id> [--dry-run] [--resume|--abandon]   Managed brains: fast-forward the checkout and sync.
gbrain sources set-strategy <id> <markdown|code|auto>
gbrain sources set-sync <id> <on|off> [--dry-run] [--expect <on|off>]
                               Include or exclude the source from automatic sync
                               (config.syncEnabled; an explicit gbrain sync --source
                               still runs). --dry-run prints the apply command, which
                               carries --expect; a moved state refuses preview_changed.
                               Off: autopilot cycles work only on pages changed since
                               the last cycle, so a file edited on disk is picked up
                               by the next explicit sync, not by the cycle.
gbrain sources host-repo <id> --dir <path> [--push-remote <url>] [--dry-run | --expect <plan_hash>]
                               Managed brains: host a new, empty source in a bare
                               repository under <path> (see below). Previews without --expect.
gbrain sources canonical <id> --to git|mixed [--dry-run [--list-out <file>] [--approval-out <file>] | --apply (--expect <hash> | --approved <file>)] [--decisions <file>] [--json]
                               Managed brains: convert a source to git-canonical or back
                               to mixed. Previews without --apply (review list, decisions
                               and approvals: docs/guides/git-canonical.md); the apply prints progress
                               on stderr and, if stopped, where it stopped
                               (docs/designs/GIT_CANONICAL_MEMORY.md#conversion).

Hosted repositories

gbrain sources host-repo gives a new, empty source its own bare repository (<path>/repo.git) that people clone and fetch from, and checks it out once as a linked worktree gbrain owns (<path>/worktree), claimed as the source's canonical worktree like any other. The source becomes git-canonical (canonical: git), so every page write, remote ones included, commits there through the door before the database indexes it. A pre-receive hook refuses every push into the repository; --push-remote names a repository of your own that the push effect publishes to once Git durability is enabled (gbrain sources writer git-durability <id> --enable).

It runs only from the trusted local CLI on the brain host. The preview says in words that the source becomes strict and prints the apply command with its --expect <plan_hash>; a plan that moved refuses preview_changed. It refuses a source that is bound or already has pages (moving existing pages into a hosted repository is conversion, a later wave), and a directory that is not empty or sits inside a checkout or another source's directory. On a hosted source the preview reports the hook and the worktree link, and the apply reinstalls a missing or changed hook. An agent runs it on a production brain only with the user's explicit yes, one source at a time.

Read-only mirror sources

A source whose Git remote is the source of truth (a code or docs repository you keep current from upstream) can be marked a read-only mirror:

gbrain sources mirror-readonly <id>

On a managed brain, sync then imports canonical metadata (a default title, type, tags kept in the brain) into the database only and never writes a file back into that checkout, so git status stays clean and the next fast-forward pull succeeds. A page write into the source (put_page, a timeline entry, a maintenance write) is stored database-only too; its receipt says storage: "database_only" with write_through.skipped: "mirror_read_only". gbrain sources list --json shows mirror_read_only per source. Undo it with gbrain sources mirror-writable <id>; files are written again from the next write on, except for pages created while the source was a mirror: those have no file in the checkout and stay database-only. Git effects of a mirror's writes (for example a forget) complete as skipped. The flag is off by default. With sync.write_through off, sync keeps that canonical metadata in the database for every source, mirror or not, and never rewrites a checkout file.

On a brain without managed persistence the flag holds the same way: page write-through (sync --full re-exports, brainstorm --save), fact fences, timeline entries, forget, the extract cycle's fence pass and owner trust repairs leave the checkout untouched and keep the write in the database (a trust repair is listed needs_review instead), reporting mirror_read_only (#7098). The dream cycle leaves such a checkout alone too: its synthesis, patterns and cycle-summary pages stay in the database, retired pages' files are kept, and no fallback file is written under .sources/ inside it (#8935).

A managed brain never pulls inside a cycle. Advance a managed checkout with one command on its owner host:

gbrain sources refresh <id>

It refuses new writes to every source that shares the checkout until queued ones finish, fast-forwards it with git merge --ff-only, then syncs each of those sources without pulling. Sources bound to one checkout (for example notes/ and docs/ of the same repository) are refreshed together, because a merge rewrites files of all of them. --dry-run previews the incoming commit; refusals are listed under worktree refresh refusals. A brain that is not managed keeps pulling inside gbrain sync --source <id>.

The autopilot cycle syncs the checkout as it is and reports upstream_refresh: "skipped_managed"; gbrain doctor (sync_freshness) says "upstream unknown" when the checkout was not fetched in the last 24 hours. See managed_pull_skipped.

Code sources: persist the sync strategy

Every sync of a source (sync --all, autopilot, the dream cycle, the MCP sync op) uses the strategy stored on the source when the call names none. Unset, that is markdown, which imports only Markdown and deletes the page of a code file that changes. Record code (code only) or auto (Markdown and code) once for a code repository:

gbrain sources set-strategy api code
gbrain sync --source api --full   # re-classify files unchanged since the last sync

gbrain sources list --json shows strategy per source (null when unset). An explicit gbrain sync --strategy still wins for that run. A source with an unfinished managed sync run refuses the change (sync_run_unfinished) until the run finishes (gbrain sync --source <id> --no-pull resumes it) or is retired (managed sync orphans): that run's frozen manifest was classified under the old strategy. Connector sources (GitHub, Google) materialize Markdown and take no strategy.

The git requirement for --path sources

A plain folder that is not inside any git repository registers as is: sync imports its files without Git, so there is no git history, and a file you remove from the folder keeps its page. A --path inside a git repository (or a subdirectory of one) needs at least one committed, tracked file under that path. gbrain sources add validates this at registration time and refuses a repository that doesn't qualify — a git init with no commit yet, or a commit made before git add — with an actionable error instead of silently registering a source that will fail (or worse, "succeed" while importing nothing) on its first gbrain sync. Fix it with:

git -C <path> init
git -C <path> add -A   # gbrain-allow-shared-tree-git: first commit of a brand-new repo nobody else writes yet
git -C <path> commit -m "initial import"
gbrain sources add <id> --path <path>

Two details that are easy to miss:

  • Files must actually be committed, not just present. The sync walker reads files through git objects, so git init alone — even followed by an empty commit (git commit --allow-empty) — isn't enough. Registration checks for real tracked content (git ls-tree HEAD scoped to the path), not just a resolvable HEAD, so this footgun is caught immediately instead of surfacing later as a sync that imports nothing. For files that are deliberately untracked or .gitignored but should still sync, use gbrain sync --include-gitignored — it forces a full filesystem walk so periodic syncs see ignored/untracked syncable content the git-object walk skips (gbrain import <dir> --include-gitignored is the one-shot import equivalent). Doctor's multi_source_drift advice names this case: a slug stuck at default whose file is untracked won't be recreated by a plain re-sync, so reach for the flag (or commit the file) before any delete step.
  • --force registers the source anyway, skipping the check. Use this if you're registering a path before an automated pipeline gets around to git init-ing it. GBrain never auto-git inits a --path source for you — it's your directory, not a gbrain-managed clone (same consent boundary as sync-time self-heal, which also never mutates a --path source without an explicit ask).

If sync ever reports a problem with the sync anchor (last_commit) — after a force-push, a history rewrite, or a from-scratch git init on a directory that was synced before — you do not need to reset anything by hand. gbrain sync detects an unreachable or non-ancestor anchor automatically and recovers: either a full reimport (anchor object missing) or a direct tree-to-tree diff against the orphaned bookmark (anchor present but rewritten), advancing the anchor to the new HEAD when it completes.

Sources in a Git subfolder

A --path source can point at a subfolder of a Git repository, for example gbrain sources add vault-notes --path ~/vault/notes where ~/vault holds the .git directory. Such a source has two possible roots for its slugs and stored file paths, called its slug-root mode:

Mode ~/vault/notes/people/alice-example.md becomes When it is chosen
source-root slug people/alice-example, stored path people/alice-example.md The default for a new subfolder source: the same slugs a plain gbrain import ~/vault/notes would give.
git-root slug notes/people/alice-example, stored path notes/people/alice-example.md When you sync with gbrain sync --repo ~/vault --src-subpath notes (you named the Git root as the base), or when the source already has pages whose slugs carry the notes/ prefix.

The mode is decided once, by the first real sync or the first coordinated page write (such as put_page) that records a new file path, and pinned in the source's configuration. A few write paths, such as saved brainstorm ideas, follow an existing pin but do not set one. Every later sync, write and reader obeys the pin, so an existing brain never has its slugs renamed. gbrain sync --dry-run works out the mode without pinning it. A source at the root of its repository, or outside Git, has only one root and no mode to choose.

Say to your agent: "Add my ~/vault/notes folder as its own source and keep its slugs relative to that folder."

Write-through. When a write creates a page in a subfolder source, gbrain writes <source path>/<slug>.md and records the stored path in the source's mode. In source-root mode the next sync of that file finds the same page instead of creating a twin. In git-root mode the recorded path adds the subfolder prefix that the unprefixed slug lacks, so the next sync can refuse that file with a slug/origin mismatch; there, add new pages as files in the checkout and sync them. Pages that already have a stored path keep writing to that file.

Older stored paths. Releases before v0.60.5.0 recorded Git-root-style paths (notes/people/alice-example.md) for pages in source-root sources. Sync accepts that form when the rest of the path names the same page, and rewrites it on the next import. No command is needed.

ambiguous_source_path. If both readings of an old stored path exist as files, for example ~/vault/notes/people/alice-example.md and ~/vault/notes/notes/people/alice-example.md, sync refuses instead of guessing. Rename or move one of the two files, commit, then run gbrain sync --source vault-notes --no-pull. The other refusal reasons are listed in write refusal reasons.

Citation format for agents

When agents receive multi-source results they MUST cite pages in [source-id:slug] form. Example:

You told me about the distillation protocol — see [wiki:topics/ai] and [gstack:plans/multi-repo] for where this came from.

The citation key is sources.id (immutable). Renaming a source via gbrain sources rename changes the display name only; existing citations keep working.

Writing to a specific source

# Pass --source explicitly
gbrain put topics/ai ... --source wiki

# Or rely on the dotfile / env / CWD match
cd ~/.gstack && gbrain put plans/multi-repo ...
# → source auto-resolves to gstack

Reads span federated sources by default. Writes require a resolved source (explicit, inferred, or default). The resolver never picks a source silently when ambiguous — it errors with a clear fix.

Unscoped writes are also guarded against landing in the wrong place. On a brain with at least one other source and more pages outside default than in it, an unscoped gbrain sync refuses (pass --source <id> to redirect it), gbrain import warns, and MCP stdio prints a once-per-process advisory when a write actually resolves to the default tier. gbrain sync --dry-run previews the run and prints the same routing guidance instead of refusing; GBRAIN_ALLOW_DEFAULT_WRITE=1 is the escape hatch when default really is the intended target. Say to your agent: "Show me what a sync would do without writing anything" — your agent runs gbrain sync --dry-run.

Durability: keep a brain repo in sync (auto-harden)

This hardening path applies to unmanaged worktrees. After activating managed writers, Git effects belong to the persistence outbox; generated legacy push helpers refuse to run rather than bypass that ownership boundary. See the concurrent-write guide. On a managed root the owner commits and pushes page writes (gbrain sources writer git-durability <source> --enable), files that are not pages are committed with gbrain git commit-push -m "<msg>" -- <paths> and ride the owner's next push, the checkout is never pulled or rebased, and the durability block below is regenerated in its managed form by that verb; see files that are not pages.

A long-lived agent that writes to a knowledge-wiki git repo needs three things to never lose work: pull before it edits, push every write, and not go stale while it sits idle. gbrain sources harden installs all of that, idempotently. The moment you add a brain repo with a token, it runs automatically:

# Clone + register a GitHub repo, then auto-harden it for durability.
# Use a fine-grained PAT scoped to just this repo.
gbrain sources add wiki --url https://github.com/you/brain-wiki.git --pat-file ~/.secrets/wiki-pat
#   → clones, then installs: local auto-push hook, scripts/brain-commit-push.sh,
#     always-on durability rules in AGENTS.md/RESOLVER.md, a 30-min pull cron,
#     and a repo-scoped credential. Verifies push works before declaring done.

# Run the same audit on an existing source any time (idempotent):
gbrain sources harden wiki --pat-file ~/.secrets/wiki-pat

# Pull on demand (the cron calls the --path form, which never opens the DB):
gbrain sources pull wiki

# Remove the durability scaffolding (also runs automatically on `sources remove`):
gbrain sources unharden wiki

What hardening guarantees:

  • Pull-first, safe in a shared checkout. Nothing rebases or stashes the working tree. A pull fetches, then, under the checkout's writer lock, fast-forwards, or for a diverged branch merges the remote commits in memory into a merge commit (gbrain: merge origin/<branch> into <branch>) and fast-forwards to it. A real conflict changes nothing and is flagged for attention (conflicts). gbrain sources pull, which the unattended 30-minute cron runs, keeps its conservative rule: a checkout with any local edit is skipped (skipped_dirty), so it never moves files under a writer, and a checkout whose writer lock is busy is skipped until the next run. An explicit scripts/brain-commit-push.sh --pull (or gbrain git commit-push --pull) advances a dirty checkout unless a dirty file is one the pull must change; then it changes nothing and its notice names the files. gbrain sources set-pull-strategy <id> <rebase|merge> is still accepted and recorded, but both strategies now integrate by merge, because the shared tree is never rebased. A force-pushed upstream is never merged: gbrain sync reports upstream_rewritten and, after the user agrees, gbrain sources reset-upstream <id> --expect-local <sha> --expect-upstream <sha> moves the checkout to the new upstream and keeps the old history under a backup ref (upstream rewritten).
  • Push is never deferred. scripts/brain-commit-push.sh "<msg>" <path> delegates to gbrain git commit-push: it takes the checkout's writer lock, commits only the named paths, merges a rejected push in memory, and refuses to report success without a push the remote accepted. Other agents' uncommitted files are never staged, stashed or touched. The post-commit hook is a best-effort background fallback (it only publishes, never pulls); the helper is the guarantee. Outcomes and what to do on each: shared-checkout git.
  • One writer at a time per checkout. Harden also sets rebase.autoStash=false, merge.autoStash=false and pull.ff=only in the checkout's local git config, so a stray git pull refuses instead of autostashing. Host daemons that commit the same checkout join through the lock contract.
  • No silent staleness. A 30-minute background pull keeps an idle session current. It runs DB-free, so it never contends with a live brain for the PGLite single-writer lock.

Flags: --no-cron skips the scheduled pull, --no-verify skips the push probe, --dry-run reports what would change, --json emits a machine report, --all hardens every source with a remote (same-account only). --no-harden on sources add opts out of auto-harden.

After upgrading GBrain, run gbrain sources harden <source-id> on each machine with an already-hardened source to refresh its local hook and scripts/brain-commit-push.sh. Upgrading the CLI alone does not update those installed scripts. Existing repo-local credentials are reused; no new token is required when that credential still works.

Security: the push automation is installed locally per machine (never committed into the repo), the token is wired per-repo (an existing credential helper is reused when present), and it never appears in the repo, the remote URL, logs, or the JSON report. For a self-hosted git server reachable only over a filesystem path, set GBRAIN_GIT_ALLOW_FILE_TRANSPORT=1 (default is HTTPS-only).

Upgrading an existing brain

gbrain upgrade runs the needed schema migrations automatically. Your existing pages all live under source_id='default'. Behavior is unchanged until you add a second source.

To add one:

gbrain sources add gstack --path ~/.gstack --federated
cd ~/.gstack && gbrain sources attach gstack && gbrain sync

Two commands. The existing default source is untouched.

Related features that build on sources

  • Session transcript ingest — gbrain transcripts (server-private: raw chat exports stay on the host machine).
  • Per-source retention — gbrain sources archive / archived / purge (soft-delete with a TTL grace window).
  • One-shot remote bootstrap — gbrain sources add <id> --url <git-url> (clone + register + auto-harden).
  • Access control across brains — the brain axis (gbrain mounts); see docs/architecture/brains-and-sources.md.