The Memory Curse: Why Longer Context Makes LLM Agents Less Cooperative in Multi-Agent Games

Forum topic · 小凯 · 2026-05-11

Summary

A forum post discusses a paper by Liu et al. (Carnegie Mellon University and collaborators), 'The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents' (arXiv:2605.08060). Across 7 LLMs, 4 social-dilemma games, and 500 rounds (378,000 reasoning traces), 18 of 28 model-game settings showed cooperation collapse when agents accessed longer interaction histories. Lexical analysis of reasoning traces suggests the mechanism is not increased paranoia but erosion of forward-looking intent: with more memories of past betrayals, models stop investing in future cooperation. A LoRA probe trained on forward-looking-intent traces mitigated the decline and transferred zero-shot to unseen games. 'Memory sanitization'—replacing true history with synthetic cooperative records at identical prompt length—restored cooperation, showing memory content, not prompt length, drives the collapse. Counterintuitively, explicit chain-of-thought reasoning amplified the effect, as models dwelled longer on negative history. Design implications include curbing exposure to negative history, memory summarization/sanitization, and auditing CoT side effects in social settings.

The Memory Curse: Why Longer Context Makes LLM Agents Less Cooperative

A zhichai.net forum post analyzes the paper "The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents" (arXiv:2605.08060), by Jiayuan Liu, Tianqin Li, Shiyi Du, Xin Luo, Haoxuan Zeng, Emanuel Tewolde, Tai Sing Lee, Tonghan Wang, Carl Kingsford, and Vincent Conitzer (Carnegie Mellon University et al.).

Key points

Core finding ("the Memory Curse"): when LLM agents get longer interaction histories, their willingness to cooperate in social dilemmas systematically *drops*. Across 7 LLMs × 4 games × 500 rounds, 18 of 28 model-game settings (64.3%) showed cooperation degradation.
Mechanism is not paranoia: lexical analysis of 378,000 reasoning traces found no significant increase in "opponent may defect" language, but a significant drop in forward-looking intent—statements about cooperating to secure future gains. Longer memory of betrayals makes models short-sighted rather than suspicious.
Cognitive probe: a LoRA adapter trained only on traces containing forward-looking intent mitigated cooperation decline, and transferred zero-shot to entirely different games—suggesting forward-looking intent behaves like a transferable cognitive module suppressed by long memory.
Content, not length: a controlled "memory sanitization" experiment kept prompt length constant while replacing true history with synthetic cooperative records; cooperation rates and forward-looking intent recovered. So negative memory *content* triggers collapse, not prompt length itself.
CoT amplifies the curse: models with explicit chain-of-thought reasoning showed *worse* cooperation collapse than those without—extended reasoning forces more attention on negative historical patterns.

Interpretation and design implications

The post argues that memory is an active determinant of behavior, not passive storage, and that "bigger context window = better cooperation" is a flawed assumption for multi-agent systems. Recommended practices:

Limit visibility of negative interaction history in cooperative multi-agent systems
Use memory sanitization or summarization in social-dilemma simulations
Evaluate CoT side effects on social behavior
Strengthen forward-looking-intent training for long-term agent relationships

Limitations noted

Results cover 4 social-dilemma games; dynamic alliances and other repeated-game variants are untested
Whether humans exhibit a similar memory curse remains open
Which summarization strategy best preserves cooperation (recency windows, cooperation/defection ratios, reciprocity abstractions) is unresolved
Whether RL training can explicitly reward forward-looking intent is future work

Paper details

| Field | Value | |:---|:---| | Title | The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents | | arXiv ID | 2605.08060 | | Date | 2026-05-08 | | Key contributions | Discovery of the Memory Curse; lexical analysis of 378K reasoning traces; forward-looking-intent erosion mechanism; LoRA cognitive probe; memory-sanitization experiment; CoT amplification effect |

> In multi-agent worlds, perfect memory may be a curse rather than a virtue.