MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
- ID: 1beb9445
- 原文链接: https://arxiv.org/abs/2608.20202
- PDF: https://arxiv.org/pdf/2608.20202v1
- 作者: Mengru Wang, Haozhe Luo, Zhenqian Xu, Zhixiang Cui, Haoming Xu, Qu Yang, Jizhan Fang, Junfeng Fang, Ningyu Zhang
- 日期: 2026-08-20
- 分类: agents
- 来源类型: arxiv
- 备注: Work in progress
- 质量评分: 4/5
- 抓取时间: 2026-08-24T23:40:00+08:00
中文导读
现有记忆 benchmark 只测信息能否被正确提取、存储和检索，忽视「忠实记录且语义相关」的记忆也可能扭曲推理与信念。MemTrapBench 定义两类认知陷阱：Reasoning Fixation（旧记忆锁死推理路径）与 Belief Distortion（记忆内容带偏信念），在两个模型家族、五个代表性记忆框架上评测，所有记忆策略都跑输无记忆设置，最强方法也跌超 10%。作者提出推理时方法 AdaptiveMem 指导模型主动规避陷阱，在缓解 MemTrapBench 陷阱的同时保持或提升标准记忆基准成绩。
Abstract (grounding)
Memory has become a key component of large language models, enabling them to retain information and learn from long-term interactions. However, existing memory benchmarks mainly evaluate whether information is correctly extracted, stored, and retrieved, while largely overlooking how retrieved memories reshape model reasoning and affect performance on the current task. We identify memory-induced cognitive traps: even faithfully recorded and semantically relevant memories can distort model reasoning or beliefs and degrade current task performance. To systematically evaluate these failure modes, we introduce MemTrapBench, which covers two forms of cognitive traps: Reasoning Fixation and Belief Distortion. Experiments across two model families and five representative memory frameworks show that MemTrapBench is challenging: all evaluated memory strategies underperform the no-memory setting, with even the strongest methods suffering drops of more than 10%. To mitigate these cognitive traps, we propose AdaptiveMem, a simple yet effective 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.
证据摘录（Obsidian 论文流水线 2026-08-24）
对所有给 agent 加长期记忆的团队，这是入库质量门控之外必须补的一层评测：记忆的「忠实」和「相关」不等于「有益」。