-Retrieve memories with a multi-round, LLM-driven search loop. An LLM plans and runs several searches over the owner's memories — keyword, semantic, hybrid, topic-cluster, and reading the surrounding conversation or a whole session — and keeps searching until it judges the evidence sufficient to answer `query`. The memories it found are returned. **The response contains memories, not an answer**: the model's own answer is discarded. Use it for questions a single search answers poorly: multi-hop, temporal, or spread across several conversations. For simple lookups, `/search` is faster and cheaper. **Request** — the same body as `/search`, except: - `method` is fixed to `\"context\"`. It can be omitted; any other value is a 422. - `min_score` is not supported (422): results are ranked by rank fusion, not by a similarity score. - `radius`, `enable_llm_rerank` and `with_readable_episode` are not accepted (422): the loop does not use them. **Response** — the same `SearchData` shape as `/search`: - The loop returns its memories in `episodes`, for a `user_id` owner only; for an `agent_id` owner `episodes`, `agent_cases` and `agent_skills` are empty. `profiles` (`include_profile`) and `unprocessed_messages` (a top-level `session_id` filter) are filled exactly as on `/search`. - `episodes[].cited` is `true` for memories the model named as the ones its answer rested on. - `episodes[].score` is a rank-fusion score, not a similarity: roughly 0.01–0.05, plus 1.0 for a cited memory, so cited memories come first. It is not comparable with `/search` scores. **Limits** — by default at most 5 LLM rounds and 45 seconds per request. Reaching a limit is not an error: the memories found so far are returned. **Latency** — several LLM calls per request, so noticeably slower than `/search`.
0 commit comments