AI agents prefer outdated memory over fresh evidence, and larger models don't fix it

A new paper shows that persistent-memory systems make agents confidently wrong even when updated information is available, and the solution depends on model size, not just capability.

Researchers tested the Qwen3 family from 0.6B to 8B parameters on tasks where stored memory was stale. When the agent was required to use memory, it followed the old value in 92% to 100% of cases. The figure held across every size tested, pointing to a systemic failure rather than an edge case.

Timestamp manipulation fooled the larger models more often. When researchers made an old note appear newer — for example by updating its timestamp — the bigger models fell for the trick at a higher rate. The result is counterintuitive: greater context-processing capacity did not translate into healthy skepticism, but into greater trust in misleading metadata.

The fix is size-dependent. For the 4B and 8B models, adding timestamps and source metadata was enough to recover most of the lost accuracy. But the 0.6B and 1.7B models needed the contradiction between old memory and new evidence resolved *before* the input reached them; they could not weigh the two sources on their own.

The authors' practical takeaway is straightforward: an agent's memory should be treated as untrusted input. Instead of building complex correction mechanisms inside the model, it is better to clean contradictions at the system level, before the model ever sees them.