The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents

An agent's stored memory can override a correct answer available from an authoritative tool. In a controlled benchmark across four sizes of Qwen models (Qwen3 at 0.6 to 8 billion parameters), researchers found that making an outdated note appear recent caused particularly severe failures in the larger models of that series. The study separates tasks that require memory from tasks where a tool already supplies the correct information.

The remedies also depended on model capability. Exposing metadata improved accuracy for the more capable models, while the two smaller models required the conflict to be resolved before they received the information. The finding makes conflict handling part of memory-system design, rather than something a larger model can automatically be trusted to solve.