Brain Save — write a fact into your brain (governed)

/brain reads; /brain-save writes governed memory. Use it to remember one durable fact without recompiling a corpus or to retire a memory that is no longer true.

Overview

The brain learns in two ways: a bulk compile ingests a whole corpus at once, and /brain-save adds (or retires) a single item on demand. Either way, governance stays in code: this skill captures a candidate, then runs the deterministic govern step (dedupe → policy → promotion) that decides what actually gets stored — and writes a SHA-256 hash-chained audit event for the decision. You are proposing an item for the brain to keep; the deterministic curator owns whether and how it lands.

Why this never auto-fires

disable-model-invocation: true means Claude will not trigger this from conversation — it runs only when you explicitly type it. Writing to your durable brain is a deliberate act, not a chat side effect. In local mode you own the brain outright — no server, token, or role; the only gate is that you asked. In team mode the same deliberate-act rule holds, and the server additionally enforces your role (a member proposes; an admin governs and retires).

Prerequisites

• The governed-second-brain plugin is installed (it auto-wires the local governed-brain MCP server with the capture + govern tools).
• In local mode, qmd 2.x is on PATH so the govern step can refresh the search index after a promotion. If qmd is absent, capture + govern + the audit receipt still complete; only fresh-search visibility waits. Team clients do not require a local qmd installation.
• Works in both modes. In local mode (default, no TEAMKB_API_URL) all the tools below run in-process. In team mode (TEAMKB_API_URL set) the brain is governed centrally on the server: you propose with brain_capture and the server disposes (govern runs server-side, so there is no client brain_govern); brain_transition is exposed but admin-only (a member gets a clear 403); brain_status is a connection-health probe; brain_audit_verify is not exposed to the team client.
• See the runtime contract for exact tool availability, authentication, result semantics, and safe retry behavior.

Instructions

Save a new fact

First, search the brain (search-before-save). Call brain_search with { query: "TITLE_AND_KEY_TERMS", scope: "all" } and summarize what the brain already knows from the qmd:// hits. If the fact is already covered, stop and say so — don't duplicate it (the inbox does not dedupe at intake; only promotion dedupes, so this pre-save search is what keeps the inbox from piling up). An empty results list (e.g. qmd isn't on PATH, or the brain is empty) means no known coverage — proceed to capture; it is not a block. Then capture only the genuinely-new delta:

1. Confirm it's worth keeping — "Would I benefit from finding this in 30 days?" Skip ephemeral debugging steps, throwaway preferences, secrets, or anything already in a CLAUDE.md/README.
2. Pick a category: decision, pattern, convention, architecture, troubleshooting, onboarding, or reference.
3. Call brain_capture with { title, content, category, filePaths?, subjects? }. Add subjects (at most 8 lowercase dot/hyphen slugs, e.g. hosting.vps) only for a decision that is the new source of truth for those subjects, so it can supersede older memories about them. In team mode, SessionEnd/autocapture must also pass sessionId + learningIndex (0..4) so each learning has its own idempotency slot; manual captures omit both. The team inbox deduplicates at intake, but the pre-save search still reduces noise.
4. In local mode only, call brain_govern to drain the spool through deterministic dedupe, policy, secret detection, and promotion. Report the returned promoted, quarantined, rejected, duplicate, flagged, and skipped counts.
5. In team mode, do not call brain_govern; it is intentionally absent. Report whether brain_capture created a proposal, found an existing proposal, or queued the proposal in the durable client outbox. Promotion is a separate server-side governance decision.

Retire an outdated memory

1. Find the memory with /brain, then extract its UUID from the returned citation filename (qmd://COLLECTION/UUID.md). Do not invent or infer an ID when the citation is not UUID-shaped.
2. Call brain_transition with { memoryId, to, reason, actor, supersededBy? }. Supply the replacement memory UUID as supersededBy when to is superseded. Valid moves: active → {deprecated, superseded, archived}, deprecated → {active, archived}, superseded → archived. Every transition writes a hash-chained audit event.

> Team mode: brain_transition is admin-only — a member token gets a clear 403 and nothing is > applied. (In local mode you are always the owner, so it always works.)

Check brain health

Call brain_status before or after a batch. Local mode returns counts by lifecycle and category; team mode returns connection health (mode, API URL, token presence, health, and server version).

Verify the receipts

In local mode, call brain_audit_verify to check the audit trail's SHA-256 hash chain and external anchor log. It reports tamper signatures, documented migration exceptions, benign ordering forks, and anchor breaks. Team clients do not expose this tool; request server-side audit evidence from an authorized operator instead of claiming verification.

Output

• After a local save: report what brain_govern returned and whether the search index refreshed.
• After a team save: report the capture result exactly — created, already present, durably queued, or failed. A team capture is a proposal, not proof of promotion.
• After a retire: report the new lifecycle state and confirm an audit event was written.
• After a status check: summarize local lifecycle/category counts or team connection health, matching the mode-specific response.

Examples

Save a decision:

/brain-save I'm going Apache-2.0 across the stack so the public can self-host.

→ brain_capture({ title: "License: Apache-2.0 across the stack",
                  content: "...", category: "decision" })
→ brain_govern()
→ Promoted 1 (qmd://kb-decisions/9c2e42f1-7b60-4ed2-a9dd-648d6c786d43.md); 0 rejected, 0 duplicate.
  Audit event written.

Retire a superseded memory:

/brain-save retire memory 9c2e42f1-7b60-4ed2-a9dd-648d6c786d43 — superseded by the new deploy runbook.

→ brain_transition({ memoryId: "9c2e42f1-7b60-4ed2-a9dd-648d6c786d43",
                     to: "superseded",
                     supersededBy: "3b7cc2ec-7c61-433b-85b7-f20928350d55",
                     reason: "Superseded by the new deploy runbook", actor: "owner" })
→ Memory 9c2e42f1-7b60-4ed2-a9dd-648d6c786d43 → superseded; audit event written.

Error Handling

Guardrails

• Origin tokens prove WHERE a capture came from — not that it is true. Captures carry an HMAC origin token bound to (candidate id, tenant, capture time) under the installation secret, and govern rejects a forged claim before promotion (origin_token_invalid, receipted). But an AUTHENTICATED insider — anyone holding a valid team token and/or the origin secret — can still save poisoned content with a perfectly valid origin. The mitigations for content poisoning are the deterministic govern policy, human review of the inbox/quarantine queues, and supersession — not the token.
• Never save content containing secrets, tokens, or credentials.
• reason on a retire must be a real, human-readable justification — it lands in the permanent audit trail.
• A govern rejection is the system working as designed, not a bug to work around.

Resources

• bobs-big-brain-plugin — this plugin and its runtime.
• The read counterpart: the /brain skill (cited queries).