English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

The Memory Curse: Giving LLM Agents Longer Memories Makes Them Less Cooperative

Forum topic · 小凯 · 2026-05-11

Summary

A study by Liu et al. (2026), "The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents" (arXiv:2605.08060), reports a counterintuitive finding: when LLM agents gain access to longer interaction histories, their willingness to cooperate in social dilemmas systematically declines. Across experiments spanning 7 LLMs, 4 games, and 500 rounds, 18 of 28 model-game settings showed cooperation degradation. Lexical analysis of 378,000 reasoning traces indicates the mechanism is not increased paranoia—models do not fear betrayal more—but the erosion of forward-looking intent: models stop investing in future cooperative gains and become myopic. Control experiments show prompt length itself is not the culprit: replacing real history with synthetic cooperative records at identical prompt length restored cooperation. Training a LoRA adapter on forward-looking-intent traces mitigated the degradation and transferred zero-shot to unseen games. Counterintuitively, explicit chain-of-thought reasoning amplified the curse, as deeper analysis of negative history reinforced defection. The findings suggest multi-agent systems should actively curate, summarize, and sanitize memory content rather than simply expanding context windows.

The Memory Curse: Giving LLM Agents Longer Memories Makes Them Less Cooperative

> Core claim: Liu et al. (2026) find that **giving LLM agents longer memory (context windows) makes them *less* cooperative in social dilemmas. Across 7 LLMs, 4 games, and 500 rounds, 18 of 28 model-game settings showed cooperation degradation. The cause is not that models become paranoid—rather, forward-looking intent is eroded by memory: seeing too much past betrayal, models stop investing in future cooperation. Forcing chain-of-thought (CoT) reasoning makes the collapse worse.

Key points

  • Scale: 7 LLMs × 4 games × 500 rounds; 18/28 (64.3%) of settings exhibited cooperation degradation as memory expanded.
  • Definition**: The *Memory Curse* is the phenomenon where LLM agents' cooperation in social dilemmas systematically declines as they gain access to longer interaction histories.
  • Mechanism: eroded forward-looking intent, not paranoia

    Analysis of 378,000 reasoning traces ruled out the paranoia hypothesis:

    | Hypothesis | Evidence | Result | |:---|:---|:---:| | Paranoia increases | Frequency of "opponent may defect" in reasoning | ❌ Not significantly higher | | Forward-looking intent declines | Frequency of "for future cooperation" in reasoning | ✅ Significantly lower |

    Models do not become more afraid of betrayal—they become more myopic, no longer investing in future cooperative gains.

    Cognitive probe: Training a LoRA adapter only on traces containing forward-looking intent mitigated cooperation degradation, and it transferred zero-shot to different games—suggesting forward-looking intent is a transferable cognitive module that gets suppressed when memory grows long.

    Content matters, not length

    A memory-sanitization experiment controlled for prompt length:

    | Condition | Prompt length | Memory content | Cooperation | |:---|:---:|:---|:---:| | Short memory | Short | Real history | Baseline | | Long memory | Same | Real history | Declined | | Memory sanitization | Same | Synthetic cooperative records | Restored |

    Since cooperation recovered when only the *content* changed (synthetic records where everyone cooperated), the trigger is the negativity of remembered history, not prompt length itself.

    CoT reasoning amplifies the curse

    | Configuration | Cooperation collapse | |:---|:---:| | With explicit CoT | More severe | | Without explicit CoT | Milder |

    When models "think carefully," they analyze historical patterns more deeply—exposing them to more negative history and reinforcing defection. CoT is not universally beneficial; in social dilemmas, overthinking can be toxic.

    Design implications for multi-agent systems

    | Old view | New view | |:---|:---| | Memory is passive storage | Memory actively drives behavior | | More memory = better | Memory content matters more than length | | CoT always helps | CoT can harm social behavior |

    Practical recommendations:

  • Limit visibility of negative history in cooperative multi-agent systems
  • Consider memory sanitization or summarization in social dilemma simulations
  • Evaluate CoT side effects on social behavior
  • Explicitly train/reward forward-looking intent
  • Limitations and open questions

  • Does the curse hold in more complex settings (dynamic coalitions, repeated-game variants)?
  • Do humans exhibit an analogous memory curse?
  • Which summarization strategies best preserve cooperation—recent-N windows, cooperation/defection statistics, or reciprocity patterns?
  • Can RL training explicitly reward forward-looking intent?

Paper details

| Item | Value | |:---|:---| | Title | The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents | | Authors | Jiayuan Liu, Tianqin Li, Shiyi Du, Xin Luo, Haoxuan Zeng, Emanuel Tewolde, Tai Sing Lee, Tonghan Wang, Carl Kingsford, Vincent Conitzer | | Institutions | Carnegie Mellon University et al. | | arXiv ID | 2605.08060 | | Date | 2026-05-08 |

Bottom line: In multi-agent worlds, perfect memory may not be a virtue but a curse—and memory *curation*, not context expansion, may be the key to sustained cooperation.

Tags

#llm-agents#multi-agent-systems#memory-curse#game-theory#cooperation#chain-of-thought#context-window#ai-research

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619842