MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
The paper argues that accurate memory can still make an LLM reason worse.
MemTrapBench tests cases where retrieved memories push models into “Reasoning Fixation” or “Belief Distortion.” Across two model families and five memory frameworks, every memory strategy trailed the no-memory setup, with the best still dropping more than 10%. The authors also propose AdaptiveMem, an inference-time instruction method meant to reduce those traps while keeping standard memory benchmark performance intact. HF Daily Papers' note
MemTrapBench tests cases where retrieved memories push models into “Reasoning Fixation” or “Belief Distortion.” Across two model families and five memory frameworks, every memory strategy trailed the no-memory setup, with the best still dropping more than 10%. The authors also propose AdaptiveMem, an inference-time instruction method meant to reduce those traps while keeping standard memory benchmark performance intact. HF Daily Papers' note
score 5