AI Agents Trust Stale Memory Over Fresh Evidence

Rohan Paul posted that AI agents can prefer stale memory to current evidence and that larger models do not reliably correct the problem. He states persistent memory can leave an agent confidently wrong even when fresh facts are available and recommends resolving stale information before it reaches the model. The tweet references a paper that tests Qwen3 models sized 0.6B to 8B on tasks involving memory.