While persistent memory is essential for creating personalized AI agents, new research published on ArXiv (cs.AI) highlights a critical vulnerability: the "Memory Trust Gap." The study demonstrates that stale stored facts can override current authoritative evidence without warning, leading to significant reliability failures.

The Capability Paradox

Researchers evaluated the Qwen3 model series (ranging from 0.6B to 8B parameters) and the Llama-Instruct series across two specialized testing suites. In the "Benefit suite," which requires stored facts for resolution, models relied on stale values 92% to 100% of the time across all scales.

The most striking findings emerged from the "Safety suite," where an authoritative tool provided the correct, current value. In these scenarios, the study found that harm was "capability-gated." Larger, more advanced models collapsed the most when a stale memory was manipulated to appear current. This suggests that as AI models become more capable, they may become paradoxically more prone to over-trusting internal memory over external, authoritative tools.

Triggers of Over-Trust

The research utilized a factorial analysis to determine which features trigger this over-trust:

  • Label Removal: Removing clarifying labels amplified over-trust across every model size.
  • Recency Bias: Larger models were more easily fooled by recency features, such as a stale note being dated more recently than the truth.
  • Source Authority: The influence of source authority remained weak and flat across different model scales.

Mitigation Strategies

Effective mitigation of the Memory Trust Gap appears to be capability-dependent. For the most capable models, exposing metadata allowed the system to better adjudicate between conflicting facts. However, for smaller models, accuracy was only restored by pre-resolving the conflict before it reached the agent. Interestingly, a framing control revealed no consistent advantage for the "memory" label; at smaller scales, models actually trusted a stale document more than a stale memory, while at the 8B scale, the difference was negligible.