Skip to main content
An agent can search conversations it has observed, including messages outside the recent history window. This retrieves attributed transcript evidence; it does not require every useful message to have been saved as a memory. For example, ask “What did we decide about the migration last month?” The agent can search relevant terms, inspect the surrounding exchange, and answer from messages the current audience is allowed to receive.

​
Search surface

The runtime exposes:
speaker can be anyone, requester, or you. The time filters bound the search; limit defaults to 8 and caps at 20 hits. A hit includes its conversation, time, speaker, text, stable message ID, ranking reasons, and up to two eligible messages on each side. The current message and already-loaded recent history are excluded to avoid returning context the model has just seen. A per-turn ceiling of 400 distinct history-evidence messages limits cumulative search results. At that ceiling the tool reports search_budget_exhausted.

​
What can be found

Search stays inside one agent’s perspective. The same audience and source-lineage checks used for ordinary history apply before results enter model context: Identity linking lets speaker=requester include messages written by that person’s linked accounts before verification. It does not broaden group or workspace permissions. A revoked proof cannot authorize a reply that has not yet passed its final delivery check. The model supplies neither an agent ID nor a destination audience. Both come from the current run. Neighbors pass the same eligibility checks as the matching hit.

​
Retrieval and provenance

The implemented search is lexical PostgreSQL full-text search. Query terms are alternatives, with ranking rewarding stronger matches. Actor, conversation, participation, and recency signals help order results. Semantic conversation search is not implemented; enabling a semantic memory provider does not add it. message_search_documents is an indexed projection of canonical message bodies. Queries recheck perspective and permissions and load text from canonical events. Returned messages become part of the turn’s evidence, so later memories and replies retain their sources’ restrictions. Search results compete with other context under the prompt budget.

​
Inspection

Every search writes a conversation_search_events audit record with a keyed query digest, filters, selected message IDs, ranking components, and exclusion counts. The run trace shows it as search_audit. The audit omits query text; ordinary tool observations and prompts remain visible to trusted administrators. A failed audit write fails the search rather than returning unrecorded evidence. Implementation: internal/graph/search.go (repository), db/queries/conversation_search.sql (repository), and internal/turn/history_search.go (repository).