When Agent Memory Becomes Stale: Provenance, Revocation, and Decision-Time Validation

An agent can retrieve an accurate historical fact and still make an incorrect present-day decision. A supplier preference may have changed, a source document may have been revoked, or a summary may combine evidence governed by different permissions. This white paper treats persistent agent memory as a mutable evidence system. It synthesizes recent memory benchmarks and established provenance and authorization work, then proposes a decision-time eligibility protocol that checks scope, lineage, validity, permissions, and conflicts before memory influences a consequential action. A synthetic procurement example exposes failure modes that recall accuracy alone misses. A reproducible evaluation design separates legitimate completion from stale evidence use, revocation exposure, and operational cost. The protocol is an engineering proposal, not an implemented system or an experimentally validated defense. Its value is a concrete contract and falsifiable test plan for teams deploying agents across changing enterprise state.

Artificial Intelligence Cloud Computing Software Engineering