The Right Memory in the Wrong Context: Verifying Retrieval Admissibility in Long-Term Agent Memory

Long-term-memory agents face risks where retrieved information is inadmissible due to policy violations or incompatible states, which standard accuracy metrics fail to detect. A new retrieval-admissibility verification framework now categorizes memory-query pairs as admissible, inadmissible, or unresolved to track these issues. Simultaneously, reports highlight storage vulnerabilities where simple phrasing can hijack an agent's memory. This exposes a critical need for auditable storage to prevent poisoned sentences from quietly shaping future sessions through rewritten beliefs.