Store a durable memory. Use for facts about the user, projects, rules, decisions you should recall later.

Hybrid recall (structural keyword + semantic vector) of the most relevant memories for a query.

Create-or-update a memory by semantic KEY (idempotent). If (namespace,key) exists it is superseded (history kept) and replaced; else created. Use for stable values an agent overwrites over time: a config value, a current decision, a user preference.

Exact key lookup: the CURRENT value stored under (namespace,key), or null. Deterministic point-read (not fuzzy recall) - reads back what upsert wrote.

Return a ready-to-inject context block of the most relevant memories for a query.

Delete a memory by id.

List folder-like namespaces and how many memories each holds.

Close a session: record a delta, the next irreversible action, and a proof artifact path.

Return a handoff pack (canon + frontiers + open loops + last checkpoint) to resume context in a new session/agent.

Anti-drift check: cash/trunk frontier set, open loops bounded, checkpoint fresh, has proof.

Set the trunk/cash/lab/parked focus (1 trunk + 1 cash + 1 lab discipline).

Digital-twin: likely stance on a situation, grounded in recorded rules and precedent.

Check a proposed action against canon/rules; flags conflicts with non-negotiable rules.

GOVERNANCE GATE: before running a tool/shell command, get a safety decision (ALLOW/WARN/HOLD/DENY/REQUIRE_CONFIRMATION/DRY_RUN_ONLY) with reasons + rollback plan. Call this BEFORE any dangerous action.

Long-session safety: interaction count vs Safe Turn Depth. When reinject_due=true, re-inject the returned canon_reminder into context - omission-rules ('never do X') decay by ~turn 10 (Security-Recall Divergence).

Pass-by-reference: get a lightweight {namespace,key,version} pointer to a memory value instead of its content (A2A courier-tax fix). Dereference with recall/find.

Optimistic-concurrency write by key: succeeds only if current version equals expected_version, else returns a conflict. Prevents lost updates when multiple agents write the same key.

Challenge a claim or proposed action against your own memory + canon BEFORE acting: flags contradictions, superseded facts, missing evidence, canon conflicts, overconfidence, dishonest omissions, irreversible actions. Returns a verdict (STOP/RECONSIDER/PROCEED WITH CAUTION/PROCEED). Call before consequential moves.

Full-system audit: memory inventory + invariants (append-only integrity, bi-temporal ordering, canon presence, dangling supersede pointers). devil=true runs the devil's advocate over every failing finding. EU-AI-Act Article-12 style record.