I Tried to Turn Agent Memory Authority Into a Scoring Formula. The Held-Out Test Changed the Claim.

A new agent-memory retrieval score adds authority, scope, specificity, action, status, and conflict terms on top of BM25.

Early tests looked promising, but a held-out evaluation showed semantic relevance alone could still pick the wrong memory.

The article now treats relevance as a tiebreaker and authority as the main retrieval goal.

It argues that agent systems should rank governance and access signals above plain textual similarity to avoid bad actions.