Fixing Memory Bias in LLM Agents

Summary & Methodology Analysis

The researchers identify that LLMs often suffer from memory-induced cognitive traps where retrieved information distorts reasoning or safety beliefs, even if that data is semantically relevant. To study this, they developed MemTrapBench, which categorizes these issues into Reasoning Fixation (encompassing cognitive bias, trauma, and task boundary issues) and Belief Distortion (specifically concerning safety). The benchmark was constructed using seed design, multi-turn dialogue generation, and a rigorous two-stage quality control process involving both automated systems and expert review. Evaluation focused on Gemini and Qwen model families using full interaction histories across five representative memory frameworks. Results consistently show that current memory integration techniques degrade performance relative to systems without memory, with top methods losing over 10 percent in accuracy. To address these failures, the authors introduced AdaptiveMem, an inference-time skill that instructs the model to identify and account for potential memory-related pitfalls before processing retrieved information. This method proved effective, showing a 14.9 percentage point improvement for LightMem on the Gemini-3-Flash-Preview model. The authors clarify that their use of the term Trauma is strictly a behavioral analogy for these failure modes and does not suggest that LLMs possess actual psychological trauma or subjective experience. The paper does not specify the exact inference latency impact or computational cost associated with the AdaptiveMem prompt-based strategy.