MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
Memory can hurt current-task reasoning even when the remembered facts are accurate and relevant.
The paper introduces MemTrapBench to test “memory-induced cognitive traps” in LLMs. It focuses on two failure modes: reasoning fixation and belief distortion. Across two model families and five memory frameworks, every tested memory strategy performed worse than using no memory, with the strongest still dropping by more than 10%. The authors also propose AdaptiveMem, an inference-time mitigation that improved results on MemTrapBench while preserving or improving standard memory-benchmark performance. ArXiv · AI/CL/LG's note
The paper introduces MemTrapBench to test “memory-induced cognitive traps” in LLMs. It focuses on two failure modes: reasoning fixation and belief distortion. Across two model families and five memory frameworks, every tested memory strategy performed worse than using no memory, with the strongest still dropping by more than 10%. The authors also propose AdaptiveMem, an inference-time mitigation that improved results on MemTrapBench while preserving or improving standard memory-benchmark performance. ArXiv · AI/CL/LG's note
score 5