MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

Summary

MemTrapBench is a new benchmark for evaluating cognitive traps in large language model (LLM) memory use. Existing memory benchmarks mainly assess whether information is correctly extracted, stored, and retrieved, but overlook how retrieved memories reshape model reasoning and affect current task performance. MemTrapBench covers two forms of cognitive traps: Reasoning Fixation and Belief Distortion. Experiments across two model families and five memory frameworks show that all evaluated memory strategies underperform the no-memory setting, with even the strongest methods suffering drops of more than 10%. To mitigate these traps, the authors propose AdaptiveMem, a simple inference-time method that instructs LLMs to avoid memory traps. AdaptiveMem mitigates cognitive traps on MemTrapBench while preserving or improving performance on standard memory benchmarks across diverse memory frameworks.