When to Remember, When to Abstain: Category-Conditioned Retention for Reliable Agent Memory (arXiv:2610.07100v1)
This paper studies conditioning write-time retention thresholds on the semantic category of an assertion (rather than using a single global confidence threshold) to improve the reliability of persistent agent memory. Evaluated on a deployed cold-start memory pipeline across 100 synthetic personas and 4,715 candidate assertions, the method lowers unsupported retentions for value assertions from 6.2% to 4.0% (~36% relative) while preserving an estimated 13 percentage points more coverage compared to a global threshold at similar retention.