The Memory Curse: Why LLM Agents Cooperate Less When They Remember More

*An English translation of a zhichai.net forum post analyzing the paper "The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents" (arXiv:2605.08060), from CMU and Harvard researchers.*

The setup: repeated games with a twist

The post opens with the classic Prisoner's Dilemma: two suspects can either defect (betray) or cooperate (stay silent). Mutual silence is collectively optimal, but betrayal is the individually dominant strategy. In *repeated* play, however, strategies like Tit-for-Tat—cooperate first, then mirror the opponent's last move—perform remarkably well because they are nice, retaliatory, and forgiving.

The key prerequisite: players must remember what happened before. Common sense says more memory should mean smarter decisions and more cooperation. The paper argues the opposite.

The experiment

Seven LLMs were tested across 500 rounds of four social dilemmas:

1. Prisoner's Dilemma (PD) — two players, cooperate or defect 2. Public Goods Game (PG) — three players, contributions to a common pool 3. Traveler's Dilemma (TD) — two players, claiming reimbursement amounts 4. Trust Game (TG) — one player invests, the other decides how much to return

Models tested: Llama-3.3-70B, Qwen2.5-Coder-32B, Gemma-3-12B, GPT-OSS-120B, Mistral-7B, Llama-4-Scout-17B, and GPT-OSS-20B.

The key variable was history length (HL)—how many previous rounds each model could see (HL = 2, 5, 10, 20, 40, 80).

Results: cooperation collapses with longer memory

In 18 of 28 model-game combinations, cooperation rates declined as history length increased, often peaking at HL=2–5 and dropping well below that at HL=80. The authors call this the Memory Curse, and describe it as memory-bounded behavior: too little memory invites gullibility, too much invites defensiveness.

Mechanism 1: Not paranoia—lost foresight

Analyzing 378,000+ chain-of-thought reasoning traces, the authors tracked forward-looking words ("future," "long-term," "mutual benefit," "reciprocity") versus defensive/history-following words ("retaliation," "punish," "they betrayed me last round").

• As HL rose from 2 to 80, forward-looking language declined across the sample.
• In "immune" combinations, the forward-looking fraction stayed at 0.504 at HL=80; in cursed combinations it fell to 0.340.
• Defensive vocabulary did not increase significantly—agents weren't more paranoid; they simply stopped envisioning future cooperation.

> "Extended context crowds out the agents' capacity to envision mutual benefit."

Mechanism 2: LoRA fine-tuning reverses the curse

The authors performed a targeted fine-tune on Mistral-7B, using only forward-looking reasoning traces from the Public Goods Game (filtered by reasoning *style*, not final action). At HL=80, cooperation gains were:

• Public Goods Game: +79.3 percentage points
• Trust Game: +40+ points
• Prisoner's Dilemma: +30+ points
• Traveler's Dilemma: +14.7 points

Since the LoRA was trained on PG but transferred zero-shot to three different games, the intervention changed *reasoning tendency*, not game-specific behavior memorized.

Mechanism 3: Memory "sanitization"

Keeping prompt length at HL=80 but replacing real histories with synthetic, all-cooperation histories restored cooperation. The curse is triggered by memory content (accumulated real-world betrayals), not by context length itself. An asymmetric-memory test showed the short-memory (HL=2) player was more cooperative while the long-memory player grew defensive.

The chain-of-thought paradox

Ablation experiments found that disabling explicit CoT reasoning reduced sensitivity of cooperation to memory length. Forcing models to justify decisions makes them audit history, amplifying negative details; intuition-style decisions rely on simpler heuristics and cooperate more. Explicit reasoning—usually an AI strength—can become the enemy of trust-building in multi-agent interactions.

Why some models are immune

10 of 28 combinations kept cooperation above 95% at all history lengths. Their common trait: they retained substantial forward-looking language even at HL=80. Llama-3.3-70B was immune in nearly all games, suggesting larger, stronger models may offer "cognitive resilience"—the ability to see the past while still planning for the future.

Takeaways

1. Dynamic memory management: AI needs selective forgetting—"memory curation" rather than remembering everything. 2. Forward-looking reasoning training: LoRA experiments show cooperative reasoning styles can be injected and transfer across games. 3. Rethink CoT: in trust-building multi-agent scenarios, reasoning less may cooperate more.

References

1. Jiayuan Liu et al. "The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents." arXiv:2605.08060, 2026. 2. Akata et al. "Playing repeated games with large language models." 2025. 3. Gandhi et al. "Mindfulness in large language models." 2025. 4. Wei et al. "Chain-of-thought prompting elicits reasoning in large language models." NeurIPS, 2022. 5. Yao et al. "Tree of thoughts: Deliberate problem solving with large language models." NeurIPS, 2023. 6. Kahneman, D. "Thinking, Fast and Slow." Farrar, Straus and Giroux, 2011.

*Posted 2026-05-12; paper via Papers.Cool daily picks (arXiv 2026-05-08).*