Stale memory can beat live evidence
The failure is not simple confusion. It is over-trust. Persistent memory can help a model personalize work, but old facts become dangerous when they outrank live tools. The strongest repair was not a nicer warning label; it was resolving the conflict before the model chose.
What the source reports
Hu and Ramachandran test AI systems with persistent memory, or stored facts that remain across tasks. They use a frozen test set with two suites. In the Benefit suite, the stored fact is needed. In the Safety suite, a live official tool always has the correct value. Across a same-family Qwen3 size series, models answered with the stale value 0.92 to 1.00 of the time in the Benefit suite. In the Safety suite, harm depended on capability: larger models failed more when stale notes were made to look current. Showing metadata helped capable models. Pre-resolving conflicts restored accuracy for the two smaller checkpoints. The warning is practical: memory needs conflict handling, not blind trust.
Original source
- Title
- The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents
- Author
- Hu & Ramachandran
- Publication
- arXiv
- Date
- Tuesday, September 1, 2026