When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors

Impact 6 arXiv cs.AI

Summary

  • LLM memory systems often delete stale facts, but hard-filtering can reduce downstream accuracy when resolver errors delete query-critical facts.
  • Targeted interventions reveal that false-drop causes significantly greater performance degradation than false-keep.
  • A deployment-matched gate achieves 0.655 accuracy compared to 0.641 for always hard-drop and 0.662 for test-informed oracle selection.
  • Memory deletion should be treated as cost-sensitive, requiring validation of false-drop risk on matched data or defaulting to label-free retention.

Why it matters

Understanding the impact of resolver errors on LLM memory filtering helps developers optimize agent reliability and prevent critical data loss during hard-filtering steps.

Who should care