MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Mengru Wang, Haozhe Luo, Zhenqian Xu, Zhixiang Cui, Haoming Xu, Qu Yang, Jizhan Fang, Junfeng Fang, Ningyu Zhang
• Venue: arXiv: cs.AI
• Published: 2026-08-20
• Collected: 2026-08-21
• URL: http://arxiv.org/abs/2608.20202v1
• PDF: https://arxiv.org/pdf/2608.20202v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

Researchers introduce a new benchmark, MemTrapBench, to evaluate how the use of memory in large language models affects their reasoning and task performance. They identify two types of cognitive traps: Reasoning Fixation and Belief Distortion. The study shows that all evaluated memory strategies underperform the no-memory setting, with drops of more than 10%. To mitigate these issues, they propose AdaptiveMem, a method to instruct LLMs to avoid memory traps.

What the Document Says

• Memory has become a key component in large language models (LLMs).
• Existing memory benchmarks evaluate the correct extraction, storage, and retrieval of information but overlook the impact of retrieved memories on model reasoning and task performance.
• The researchers identify two forms of cognitive traps: Reasoning Fixation and Belief Distortion.
• MemTrapBench is introduced to systematically evaluate these failure modes across two model families and five memory frameworks.
• Experiments show that MemTrapBench is challenging, with all evaluated memory strategies underperforming the no-memory setting, suffering drops of more than 10%.
• AdaptiveMem is proposed as a method to instruct LLMs to avoid memory traps and mitigate cognitive traps on MemTrapBench while preserving or improving performance on standard memory benchmarks.

Key Concepts Mentioned

• Memory
• Large Language Models (LLM)
• Cognitive Traps
• Reasoning Fixation
• Belief Distortion
• MemTrapBench
• AdaptiveMem

Claims Made by the Document

• Existing memory benchmarks have limitations.
• Memory-induced cognitive traps can distort model reasoning or beliefs and degrade current task performance.
• All evaluated memory strategies underperform the no-memory setting, with drops of more than 10% on MemTrapBench.
• AdaptiveMem mitigates cognitive traps and improves or preserves performance on standard memory benchmarks.

Evidence Used

• The text describes the role of memory in LLMs and the limitations of existing memory benchmarks.
• The researchers introduce MemTrapBench and explain its purpose to evaluate cognitive traps in LLM memory use.
• Experimental results demonstrate the performance of various memory strategies on MemTrapBench.
• The text describes AdaptiveMem as a method to instruct LLMs to avoid memory traps and its effectiveness.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.