MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

TL;DR - MemTrapBench evaluates how relevant, correctly retrieved memories can impair LLM reasoning through fixation and belief distortion. Across tested model families and memory frameworks, every memory strategy performed worse than using no memory, while the proposed AdaptiveMem mitigation reduced these failures.

• Tests Reasoning Fixation and Belief Distortion rather than only memory storage and retrieval accuracy.
• Covers two model families and five representative memory frameworks.
• Even the strongest evaluated memory methods suffered performance drops exceeding 10% versus the no-memory setting.
• AdaptiveMem uses inference-time instructions to avoid memory traps while preserving or improving standard memory-benchmark performance.