Signal MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

Summary

Memory has come to serve as a core capability that lets large language models hold onto information and improve through interactions that unfold over time. Yet most existing memory benchmarks only check whether information was correctly pulled out, stored, and retrieved, while largely leaving aside how a retrieved memory actually reshapes the model's reasoning. The authors point out that even memories that are accurately recorded and genuinely relevant to the task can distort a model's reasoning or beliefs and hurt its performance on the task at hand, a pattern they label memory-induced cognitive traps. To evaluate this systematically they introduce MemTrapBench, which covers two trap types they call reasoning fixation and belief distortion. Tests across two model families and five representative memory frameworks found that every memory strategy performed worse than using no memory at all, with even the best method losing more than ten percent. As a countermeasure the authors propose AdaptiveMem, an inference-time technique that tells the model to steer clear of these traps, and report that it keeps or even improves performance on standard memory benchmarks.