---
title: "MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.20202v1"
permalink: "https://feed.astralor.com/items/itm_bb88ef095c9b7c00e38262d8"
published: "2026-08-20T16:00:17.000Z"
score: 76
channel: "ai"
tags: ["AdaptiveMem","LLM","MemTrapBench","Cognitive Trap","Memory Benchmark"]
---

# MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

## AI commentary

> 这篇论文最值得看的是它揭示了“忠实记忆也能坑模型”的盲区，提醒做记忆系统的人别只盯着存取精度。对于正在构建长期 Agent 记忆的工程师和研究者来说，MemTrapBench 和 AdaptiveMem 都是绕不开的参考。

## At a glance

Existing memory benchmarks for large language models mainly evaluate whether information is correctly extracted, stored, and retrieved, but largely overlook how retrieved memories reshape model reasoning and affect current-task performance. MemTrapBench focuses on two forms of memory-induced cognitive traps: Reasoning Fixation and Belief Distortion. Experiments show that all evaluated memory strategies underperform the no-memory setting on MemTrapBench, with even the strongest methods suffering drops of more than 10%. The authors propose AdaptiveMem, a simple yet effective inference-time method that mitigates cognitive traps while preserving or improving performance on standard memory benchmarks.

## Key points

- MemTrapBench covers two types of cognitive traps: Reasoning Fixation and Belief Distortion, testing how memories negatively affect current reasoning.
- Experiments show that all existing memory frameworks underperform the no-memory baseline on MemTrapBench, with even the strongest dropping over 10%.
- AdaptiveMem uses inference-time instructions to steer models away from memory traps, mitigating harm without sacrificing standard memory benchmarks.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_bb88ef095c9b7c00e38262d8