Interface RollingSummaryMemoryUpdate

Optional organization context (multi-tenant / org-scoped memory).

Optional routing mode (customFlags.mode, persona id, etc.).

Compaction profile id used for this update (if profile routing is enabled).

Effective long-term memory policy for this conversation at the time of compaction. Implementations should respect this (e.g., allow per-conversation opt-out).

The rolling summary markdown (human-readable).

The structured memory JSON (memory_json) emitted by the compactor.

Timestamp up to which messages are considered summarized.

When this summary snapshot was updated.

With a learning layer: the turn's scope ids, from the GMI's learner session. The sink files the user, persona and conversation targets under them and records the upsert through recordSelfChange; without them it files under the plain ids and records nothing, as with no layer.

With turnScopes: the learner session's recorder, bound to the turn. A sink that files under the turn's scope ids records each upsert through it as one memory write, or files nothing.

With turnScopes: the explicit user ids whose messages this compaction and the conversation's earlier ones summarized. Each trace the sink files records them, and the recorder notes a session row for each with the write, so forgetUser for any of them finds the trace, in the session's scopes and in its learner's own.