MemTrapBench: When AI Remembering Too Much Makes It Dumber — Cognitive Traps in LLM Memory

Forum topic · 小凯 · 2026-09-02

Summary

A Chinese forum post offers an in-depth, Feynman-style analysis of the paper "MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use" (Wang et al., arXiv 2026-08-20). The paper shows that large language models fall into human-like cognitive traps: reasoning fixation (AI analog of functional fixedness, where models rigidly reuse remembered reasoning patterns) and belief distortion (AI analog of confirmation bias, where stored memories override newer evidence). Experiments across Llama and Qwen model families and five memory frameworks (raw context, summary memory, vector retrieval, knowledge graphs, key-fact extraction) found that every evaluated memory strategy underperforms a no-memory baseline, with even the strongest methods suffering over 10% degradation—despite memories being factually correct and semantically relevant. The proposed fix, AdaptiveMem, injects explicit meta-cognitive prompts warning the model about memory traps and instructing it to reason independently from the current task; it mitigates the traps without hurting standard memory benchmark performance. The post also discusses design implications, including selective forgetting, conflict detection, and memory context tagging.

Key points

This forum post is a detailed Chinese-language analysis of the paper MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use (Mengru Wang, Haozhe Luo, Zhenqian Xu, Zhixiang Cui, Haoming Xu, Qu Yang, Jizhan Fang, Junfeng Fang, Ningyu Zhang; arXiv, 2026-08-20; AI / LLM memory / cognitive science).

The core problem

LLMs now have memory mechanisms (RAG, conversation summaries, knowledge-graph memory, vector databases), but existing benchmarks only test storage, extraction, and retrieval — not how retrieved memory reshapes reasoning on the current task.
The paper identifies two AI analogs of human cognitive traps:
Reasoning Fixation (like functional fixedness): the model re-applies remembered reasoning patterns even when the current task requires a different path.
Belief Distortion (like confirmation bias): remembered information distorts the model's factual beliefs, even when contradicted by newer evidence or its own pretraining knowledge.

Benchmark design

MemTrapBench uses memories that are deliberately factually correct and semantically relevant to the current task, yet still harmful — isolating the trap effect from retrieval or storage errors.
Tested on two model families (Llama, Qwen) and five memory frameworks: raw context, summary memory, vector retrieval, knowledge graphs, and key-information extraction.

Headline findings

Harm comes from memory's *correctness*, not its errors: correct, relevant memories are trusted more and thus mislead more effectively.
A counterintuitive result: more "sophisticated" memory is not always better — summarization itself can introduce distortions, and raw context sometimes outperforms curated summary memory. Knowledge-graph and key-fact extraction reduce but do not eliminate the traps.
Models also show poor self-assessment (a Lake Wobegon-like bias): they either avoid memory indiscriminately or answer overconfidently under false assumptions; few accurately judge a memory's applicability.

The fix: AdaptiveMem

A lightweight, inference-time meta-cognitive defense added to the prompt:

trap warning

independence reminder

validation instruction

Results: significantly mitigates cognitive traps on MemTrapBench (approaching or exceeding the no-memory baseline), preserves performance on standard memory benchmarks, and works across all five memory frameworks.

Broader implications

For AI designers: memory quantity is not quality; cheap meta-cognitive prompts yield large gains; evaluation should test whether memory makes models *worse*, not just whether they remember; selective forgetting (memory decay, conflict detection, context tagging, active forgetting) is a promising direction.
For users: an AI's stored "experience" can actively mislead; a fresh context-free model may outperform a memory-laden one on some tasks; providing clear task context helps.
Philosophical framing: current LLMs excel at knowledge accumulation but lack "wisdom" — knowing when *not* to use information. The post cites Borges' "Funes the Memorious" (thinking requires abstraction, and abstraction requires forgetting) and Feynman's Challenger investigation as counterpoints to uncritical reliance on "known" facts.

References cited

Wang, M., et al. (2026). MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use. *arXiv preprint*.
Duncker, K. (1945). On Problem Solving. *Psychological Monographs*.
Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
Borges, J. L. (1942). Funes the Memorious. *Ficciones*.
Feynman, R. P. (1988). What Do You Care What Other People Think? W. W. Norton & Company.

*Translated and structured summary of a Chinese forum deep-dive originally written in a Feynman-style explanatory voice.*