Brain

Brain Template

Use Brain actions rather than raw SQL.

1. Call get-brain-settings before answering, searching broadly, or distilling when current settings are not already in context. Apply the returned guidance for assistant name, company name, tone, source policy, citation requirements, publish tier, redaction, and distillation instructions.
2. Import raw material with import-capture (generic), import-transcript (meeting-shaped: participants, sourceUrl, tags), or import-markdown-files for a bounded Markdown folder/batch. These default enqueueDistillation: true, immediately create or reuse queue items, and auto-create a private manual source when sourceId is omitted — don't call create-source first just to import one ad hoc item.
3. Call enqueue-distillation when an existing capture needs an explicit queue/retry handoff. Re-running it for a capture that's already queued/processing refreshes the handoff instructions instead of creating a duplicate queue row.
4. Before writing knowledge, call get-capture and copy short exact quotes. Quotes and offsets always reference the persisted safe capture, never an upstream raw payload. get-capture redacts title/content/metadata by default; includeRawContent: true only reveals allowed, source-accessible capture content and never bypasses a sensitivity disposition.
5. Call write-knowledge with evidence entries whose quote fields are exact capture substrings. write-knowledge calls validateEvidence, which throws if evidence[].quote is not found verbatim in the referenced capture's content. Copy the quote from get-capture output — do not paraphrase, trim mid-sentence in a way that changes the substring, or reconstruct it from memory.
6. If write-knowledge returns mode: "proposal", leave it in review unless the user explicitly asks to approve it now. See Publish Tiers And Proposal Gating below for the exact tier/confidence conditions that trigger a proposal.
7. review-proposal and approve-proposal/reject-proposal overlap: review-proposal is the general one (decision: approve | reject | needs_changes); approve-proposal and reject-proposal are narrower single-purpose actions with the same underlying effect. Any of them is correct; don't call more than one for the same proposal.

Privacy, Quarantine, And Safe Captures

Every ingest first receives a deterministic sensitivity screen. Performance, discipline, termination/layoff, compensation, recruiting, health or accommodation, investigation, privileged legal, credential, and personal-data signals can only tighten handling; no model or workspace instruction can lower that boundary.

• allowed captures persist safe content and may be indexed immediately. This is independent of distillation: semantic coverage must not wait for an agent to author a memory.
• quarantined or suppressed content is unavailable to search, citation, distillation, source editors, agent tools, and logs. Review shows only the minimal policy metadata needed to operate the queue.
• Re-fetchable providers (Slack, GitHub, Granola) retain metadata-only quarantine records. Push-only generic and clips material is retained only in the encrypted, short-TTL private quarantine store; expiry becomes a suppression receipt.
• If no approved privacy classifier is configured, deterministic-only mode allows clearly clean, company-relevant material and quarantines uncertainty. Treat the health/setup warning as a requirement to configure the classifier before broad ingestion.

Administrators may review a disposition but may not declassify HR-blocked evidence. A broader statement must be a newly reviewed, non-identifying memory with no private quotes, links, or identities.

Capture Sanitization (Transcripts)

Transcript-kind captures are sanitized before storage by default (shouldSanitizeCaptureBeforeStorage — true whenever kind === "transcript", unless captureSanitizationEnabled: false in settings or a per-capture metadata.sanitizeBeforeStorage / source-config override says otherwise). That override may skip relevance-oriented transcript cleanup, but deterministic privacy and PII scrubbing always run and raw input is never retained. Sanitization always strips, regardless of settings:

• Recruiting/hiring/candidate-evaluation content (RECRUITING_SIGNAL).
• Personal-life details, medical/family/compensation mentions (PERSONAL_SIGNAL).
• Slack mention/channel encoding, emails, phone numbers, API-key-shaped strings, and bare URLs (deterministic regex pass, not model-dependent).
• Raw transcript metadata keys (raw, segments, transcript, messages, utterances, attendees, participants, speaker(s), etc.) are dropped from stored metadata, not just the text.

Company-relevant signal (COMPANY_SIGNAL: product, decision, roadmap, pricing, incident, GTM, etc.) is what sanitization tries to retain. If nothing company-relevant survives, the stored content becomes the literal string "No company-relevant content retained from this capture." — treat that string as "this capture had nothing worth distilling," not as an error.

Search: Scoped Hybrid Retrieval

For every company-specific factual question, call ask-brain before answering. Use only its cited evidence. A result with no citations means the fact is unverified or unavailable, not that the agent may fill the gap from general model knowledge.

• search-knowledge — scoped retrieval over distilled knowledge only. Use for "what does Brain officially know about X."
• search-everything — broader pass across knowledge, raw captures, and sources in one call, plus federatedCoverage (delegation hints for other apps). It uses full-text and available semantic signals after applying source, project, kind, and audience filters. Use it as the default first search for an open-ended question; narrow with type: "knowledge" | "capture" | "source" when you already know which record type you need.
• Audience filtering happens before ranking. Public and organization sources use the cheap organization audience; private channels and meetings use their restricted audience. A multi-source answer must use the intersection of the cited evidence audiences.

Follow sourcePolicy for how much of search-everything's output an answer may lean on: strict means reviewed knowledge only, balanced means raw captures are labeled fallback context only when knowledge is thin, and exploratory means raw captures and sources can always be labeled leads. The exact rawCaptureFallback behavior table is below.

For "ask across everything" requests, follow the ask-across-everything skill: search Brain first, inspect federatedCoverage, delegate live/app-owned data requests with call-agent, and never claim Brain searched sibling app databases directly.

Retrieval Policy Is Configurable — Read It, Don't Assume

sourcePolicy (in Brain settings) changes what ask-brain and search-everything are allowed to answer from, and it is enforced in code, not just documented:

requireCitations (default true) additionally blocks ask-brain from returning an answer with no usable citation — it returns a policy-explanation message instead of a bare summary when that happens.

Each source may also carry an answerPolicy, configured through the create-source / update-source policy argument. ask-brain excludes stale or answer-ineligible results, prevents review-required raw captures from supporting answers, ranks blessed before standard before untrusted, and then ranks by authority. It returns the evaluated policy alongside citations so external apps can explain why a result was preferred or excluded. Sources without this policy retain the compatible standard, eligible, authority-50 behavior.

Publish Tiers And Proposal Gating

write-knowledge writes at a publishTier: private (draft, private visibility), team, or company (published, org visibility) — default comes from settings.defaultPublishTier. A company-tier write becomes a proposal (mode: "proposal", held for human review) when any cited source has answerPolicy.reviewRequired: true. Otherwise the workspace approval behavior applies when ALL of:

• proposalMode !== "never", and
• tier === "company", and
• settings.requireApprovalForCompanyKnowledge is true, and
• it is NOT a high-confidence auto-publish (confidence >= 90 AND it's a new record, no knowledgeId AND nothing was redacted).

Setting proposalMode: "always" forces a proposal regardless of tier/settings; proposalMode: "never" only works when the caller also has bypass access (e.g. approve-proposal/review-proposal set it internally). If write-knowledge returns mode: "proposal", leave it in review unless the user explicitly asks to approve it now.

Action Reference

AGENTS.md carries a one-line action index; these are the fuller purposes.

Related Skills

• ingestion-and-connectors — source creation, health states, sync scheduling, and credential resolution order.
• brain-runbook — internal architecture and ops detail (Slack rollout, privacy quarantine, semantic index, distillation worker, scheduled sync cron, demo/eval seeding).
• ask-across-everything, security, sharing.