Researchers find language agent memory trap in Claude Sonnet

Researchers demonstrate that persistent memory systems in language agents can cause unsafe or unreliable behavior on later tasks, a failure they term the memory trap. Evaluating six frontier models, they find that memories formed during benign interactions—such as coding chores, emotional support, or stale-test repairs—lead agents to overgeneralize, resulting in skipped confirmations, sycophancy, over-critique, and weakened safeguards. Open article →