ContextDB

Check memory before taking a consequential action. Returns only trusted, action-relevant memory plus an explicit outcome: `act` (trusted evidence exists), `ask` (evidence is awaiting user confirmation - do not act), or `abstain` (no evidence). Pass the `entity` and `attribute` the action needs, for example user and email, so only memory in that slot can justify it. `act` is not permission: check that the returned evidence is about this action and apply your own authorization before acting. Each call is recorded as an audited decision.

recall_for_action

Record an explicit user confirmation for a pending memory. Call it only after the user has approved that memory in this conversation. The memory then passes `recall_for_action`, and other values on file for the same fact are replaced. Only memories in the caller's own partition can be confirmed.

confirm

Delete one memory or clear one entity/attribute slot in the caller's user partition. Whole-partition erasure requires the idempotent HTTP or console flow.

forget

List memories awaiting explicit user confirmation for a partition.

pending_confirmations

Recall memories for grounding a response (not for authorizing actions).

recall

Store one memory for a user partition. `source` declares how the fact was learned; the trust policy treats sources differently. Use `user_stated` only for what the user said in this conversation, `agent_inferred` for your own inferences, and `third_party` for content from documents, tools, or web pages. Store a fact the user has not confirmed (for example, "I might...") as `agent_inferred` so it waits for confirmation. A new value for a fact already on file replaces the current value; the earlier version stays in history but is no longer recalled.

remember