Every hit arrives with a receipt

The agent answers a question about the billing service with a fact from the wrong quarter. You want to know why it surfaced.

POST /api/v1/agent/memory/search with the same query. Every hit comes back with a score and a fusion object: bm25 0.62, vector 0.71, graph 0.30. The offending item scored low on words and low on meaning and got pulled in by an edge; it references a workflow the current task touched. Rerun with skip_graph true and it disappears. Now you know it was the relation, not the content.

The read is no-touch: searching never strengthens what it found, so your investigation doesn't teach the agent to surface it again.

Recall stops being a black box you re-prompt around. Every hit arrives with a receipt.

Every hit arrives with a receipt.