Verb-shaped agent retrieval — active or deep mode

Explicit session id, honored on the string-scope path

Query text (at most 4096 characters)

Retrieval mode

Override default top_k

Drop hits scoring below this. Without it, recall returns its full top_k nearest neighbours whatever their relevance — so escalating to the deep tier 'when active recall comes back empty' never fires against a non-empty store. Escalate on this threshold, not on an empty result.

Attach a chronological timeline string to the response

Agent identifier for retrieval scoping; falls back to default-agent if unset

Ask for probable_answer in the response (default true). False skips it; the list is the same either way.

Relative confidence, NOT a calibrated probability. It is a sigmoid over the reranker's top-1 logit whose constants were never fitted (see services/retrieval_pipeline.py). It orders results correctly; its absolute value means nothing, so do not threshold on it — use min_score.

Set when the active-memory tier failed during this call; hits then come only from the remaining tier(s) (degraded, not empty-by-design).

Set when the deep (graph) tier failed during this call; hits then come only from the active tier (degraded, not empty-by-design).

Why probable_answer is null (null when it is set): no_memory_answers = no memory in the list answers the question; unavailable = it could not be worked out this time; not_enabled = not requested or not enabled for this deployment.