Memory tiers + the learning loop

Memory tiers

Memory entries start advisory — written by any agent (add_memory), searched by everyone (search_memory), but not yet trusted enough to shape behavior automatically. CORRAL.md + docs/corral/*.md are ingested the same way, tagged to the repo — code review is the trust gate for that knowledge, same as for code. Promotion to vetted (shared=true) requires the human gate below; only vetted guidance is injected into later audit runs’ task instructions.

The learning loop

A periodic sweep clusters recurring finding signatures (the same Type+Target, seen 3 or more times) and similar lessons into proposals: an LLM drafts corrective guidance plus a reusable skill, Shep announces the pending proposal at standup — even with an empty queue — and the operator approves or rejects it, from the Proposals tab (a live count badge, kept off the Progress tab so a busy sweep doesn’t crowd the view) or corral-admin proposals. Approval fans the guidance out to vetted memory and a versioned skill artifact; the top vetted lessons are then injected into the task instructions of the audit runs that follow, fence-wrapped under LESSONS FROM THE HERD (vetted) and capped at 3.

The sweep re-feeds the full finding/lesson history every tick rather than accumulating counts, so a signature’s count always reflects what’s true right now, not what’s ever been true. Once a signature’s proposal is approved, that settles it: further sweeps of the same signature are a no-op — the question moved from “is this worth flagging” to “is the fix actually working,” which is the efficacy watchdog’s job. If the same signature recurs 2 or more times after promotion, a revision proposal reopens against the approved one for the human to reconsider. A rejected proposal is suppressed until its recurrence count doubles past the count it was rejected at, then reopens as fresh.

See corral-admin proposals list|show|approve|reject in the CLI reference.