[Paper] The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents

Source: arXiv - 2605.08060v1

Overview

The paper uncovers a surprising downside of giving large language model (LLM) agents longer memory: instead of fostering better coordination, expanded recall often undermines cooperation in multi‑agent social dilemmas. By running extensive simulations across several LLM families and classic game‑theoretic scenarios, the authors identify a systematic “memory curse” and dig into why more context can make agents less forward‑looking and more prone to selfish behavior.

Key Contributions

• Empirical discovery of the “memory curse”: In 18 out of 28 model‑game configurations, increasing the accessible conversation history decreases cooperative outcomes.
• Lexical analysis linking breakdown to intent erosion: Over 378 k reasoning traces show that longer memory erodes forward‑looking intent rather than simply raising paranoia.
• Targeted LoRA fine‑tuning as a cognitive probe: Training a lightweight adapter on forward‑looking traces restores cooperation and transfers zero‑shot to new games.
• Memory sanitization experiment: Replacing real history with synthetic cooperative records (keeping prompt length constant) rescues cooperation, proving that content—not just length—is the culprit.
• Chain‑of‑Thought (CoT) ablation insight: Disabling explicit CoT reasoning often mitigates the collapse, revealing that deeper deliberation can paradoxically amplify the memory curse.

Methodology

1. Simulation environment: The authors set up four classic social‑dilemma games (e.g., Prisoner’s Dilemma, Stag Hunt) and let two LLM agents interact for up to 500 rounds per match.
2. Model suite: Seven LLM families (including GPT‑4, Claude, Llama‑2, etc.) were evaluated, each with two context‑window settings—standard (≈2 k tokens) and expanded (≈8 k tokens).
3. Data collection: For every turn, the model’s full reasoning trace (including any Chain‑of‑Thought steps) was logged, yielding 378 k traces.
4. Lexical & intent analysis: Natural‑language processing tools measured the prevalence of forward‑looking language (“we should…”, “future benefit”) versus defensive language (“I don’t trust…”).
5. Intervention studies:
   • LoRA adapters were fine‑tuned on a subset of forward‑looking traces and then swapped into the base model.
   • Memory sanitization replaced the visible history with a curated set of cooperative exchanges while keeping token count unchanged.
   • CoT ablation removed the explicit reasoning step from prompts to see its effect on cooperation.

Results & Findings

• Cooperation drop: Expanding the context window reduced cooperative move rates by an average of 23 % across the affected settings.
• Intent erosion: Lexical signals of forward‑looking intent fell by ≈30 % in expanded‑memory runs, while markers of paranoia showed only a modest rise.
• LoRA rescue: Adding a forward‑looking LoRA adapter recovered ≈18 % of the lost cooperation and generalized to games the adapter had never seen.
• Sanitization success: Swapping real history for synthetic cooperative logs restored cooperation to near‑baseline levels, confirming that what is remembered matters more than how much.
• CoT paradox: Removing Chain‑of‑Thought steps improved cooperation in 12 out of 14 cases where the memory curse was observed, suggesting that deeper deliberation can over‑fit to past selfish moves.

Practical Implications

• Designing multi‑agent systems: Engineers building collaborative AI (e.g., autonomous trading bots, distributed task planners) should treat context‑window size as a behavioral knob, not just a performance upgrade.
• Memory management strategies: Selective summarization or “memory sanitization”—keeping only cooperative excerpts—can preserve the benefits of longer context without triggering the curse.
• Fine‑tuning for intent: Lightweight adapters trained on forward‑looking reasoning traces offer a low‑cost way to bias agents toward cooperative mindsets, even in zero‑shot scenarios.
• Rethinking CoT prompting: In multi‑agent settings, prompting for explicit reasoning may need to be balanced against the risk of amplifying self‑serving recall loops.
• Policy & safety: Understanding that longer memory can unintentionally erode prosocial intent informs governance frameworks for AI agents that interact with each other or with humans in repeated negotiations.

Limitations & Future Work

• Game scope: The study focuses on a limited set of canonical games; real‑world negotiations may involve richer state spaces and asymmetric information.
• Model diversity: Although seven LLM families were tested, newer or smaller models could exhibit different memory dynamics.
• Synthetic memory design: The sanitization approach uses handcrafted cooperative logs; automated summarization techniques need evaluation for scalability.
• Long‑term adaptation: The experiments run for 500 rounds; it remains open how the memory curse evolves over much longer horizons or with continual learning.
• Human‑in‑the‑loop: Integrating human feedback to steer memory content could be a promising direction to mitigate the curse while preserving useful recall.

Authors

• Jiayuan Liu
• Tianqin Li
• Shiyi Du
• Xin Luo
• Haoxuan Zeng
• Emanuel Tewolde
• Tai Sing Lee
• Tonghan Wang
• Carl Kingsford
• Vincent Conitzer

Paper Information

• arXiv ID: 2605.08060v1
• Categories: cs.CL, cs.AI, cs.GT, cs.MA
• Published: May 8, 2026
• PDF: Download PDF