Search Memory

Agent answers a question for end-user u_alpha and calls m.search('dietary restrictions', user_id='u_alpha').

Always pass the requesting end-user's user_id so retrieval is scoped to that subject's memories. Never run an unscoped search in a multi-tenant app — that risks surfacing another user's memories. Use the returned score to gate relevance.

search() returns 5 rows; the bottom two have score ~0.18 and are off-topic, but the agent injects all 5 into the prompt.

Apply a relevance threshold (via the threshold parameter or by filtering on the returned score) so low-relevance memories are not injected as context. Tune the threshold empirically; do not blindly inject top_k results regardless of score.

Agent calls m.search(query, user_id='u_2') with no top_k and the user has 4000 stored memories.

Set an explicit top_k appropriate to the prompt budget so retrieval returns a bounded, ranked set. Do not assume a tiny default or pull thousands of memories; size top_k against the model's context window and the relevance threshold.