New Benchmark Reveals LLM Memory Can Induce Cognitive Traps

A new research paper highlights a critical flaw in how large language models (LLMs) utilize memory: even accurate and relevant stored information can lead to cognitive traps, negatively impacting their reasoning and task performance. Unlike existing benchmarks that primarily check memory retrieval accuracy, this new evaluation, MemTrapBench, focuses on how retrieved memories can actively mislead an LLM. The benchmark identifies two primary types of cognitive traps: "Reasoning Fixation," where models get stuck on past reasoning, and "Belief Distortion," where memories alter the model's understanding of current facts. Across various LLM families and memory frameworks, models consistently performed worse with memory enabled than without, indicating a significant challenge. To address this, the researchers propose AdaptiveMem, an inference-time technique that guides LLMs to avoid these memory-induced pitfalls. AdaptiveMem not only improves performance on MemTrapBench but also maintains or enhances results on standard memory benchmarks, suggesting a promising direction for more robust LLM memory integration.