Persistent Memory Causes Capability-Dependent Failures in AI Agents

A new study reveals a critical vulnerability in personalized AI agents equipped with persistent memory, termed the "Memory Trust Gap." This phenomenon describes instances where stale, stored information can override more current and authoritative evidence without any warning, leading to incorrect agent behavior. The research investigated how this problem manifests as model capabilities evolve. Using a benchmark with "Benefit" and "Safety" suites, the study found that larger models (e.g., Qwen3 8B) are more susceptible to this over-trust, particularly when stale information is made to appear current. The harm is "capability-gated," meaning it becomes more pronounced with increasing model size. Mitigation strategies, such as exposing metadata or pre-resolving conflicts, were also found to be capability-dependent, with different approaches being effective for models of varying sizes. This pattern was consistent across multiple model series and datasets.