Brain Memorybrain memory
Get started

/brain:verify

Writes from untrusted sources — tool output, external content, or text that looks instruction-shaped — are held in a pending-verification state (see Provenance & Trust). verify is how they get resolved. Sensitive-topic writes are held the same way (reason sensitive_opt_out) until you opt in or approve them one by one. Resolution is always the user's call; the agent never approves on its own judgment.

info

"Verify" is distinct from the spaced-repetition review queue (due_for_review). Verification is a one-time trust decision on a quarantined write; review is ongoing reinforcement scheduling.

Usage

brain verify list                    # pending memories (JSON)
brain verify show <id>               # full frontmatter + body for inspection
brain verify approve <id> [<id>…]    # clear the flag, mark vetted
brain verify reject  <id> [<id>…]    # archive (recoverable)
brain verify requeue <id> [<id>…]    # undo an approve: back to pending

What approval, rejection and requeue do

  • Approve clears the quarantine flag and marks the memory vetted. Its origin, write-time caps, and recall trust weighting all stay — approval means "a human looked; this is not an injection," not "this is now user-asserted." To fully trust it, re-store it with origin: user or pin it explicitly.
  • Reject archives the memory to ~/.brain/_archived/ (recoverable). Pass --force only if the CLI refuses a high-salience entry and you're sure.
  • Requeue is the undo for an approve. It puts the memory back into pending verification with reason requeued and clears vetted — which also withdraws per-item consent, so a sensitive memory goes back to hidden. It refuses a pinned memory (unpin it first), and it does not take back a replacement the approval released: the older memory stays demoted, and the CLI reports it under supersession_kept.

Every approve, reject and requeue is written to the append-only audit log.

warning

Approve the ids you reviewed, not the queue. Other sessions and agents write to the same brain, so the queue can grow between the moment you looked at it and the moment an agent acts. An agent that re-lists and approves "everything pending" can sweep in a memory nobody saw — requeue is how you put it back.

Sensitive-topic writes

A memory classified sensitive — health, race, ethnicity, religious beliefs, politics, gender identity or sexual orientation (see Sensitivity) — is quarantined with reason sensitive_opt_out unless sensitive_topics: true is set in ~/.brain/config.json. Unlike an ordinary pending write, it is hidden from recall and session start whatever quarantine_mode says, so nothing surfaces while it waits.

brain verify list shows the reason. Approving is per-item consent: it lifts the exclusion for that memory only, and its receipt keeps the ⚠ sensitive marker. Setting sensitive_topics: true opts in for future writes; it is never retroactive — it does not approve memories already held, and never recovers a refused blocked write.

Related