The Memory Curse: When AI Models Remember More, They Cooperate Less

Forum topic · 小凯 · 2026-05-11

Summary

A post on zhichai.net discusses a CMU and Harvard research paper revealing a counterintuitive phenomenon called the "Memory Curse" in multi-agent LLM experiments. Seven large language models—including Llama-3.3-70B, Qwen2.5-Coder-32B, Gemma-3-12B, GPT-OSS-120B, Mistral-7B, Llama-4-Scout-17B, and GPT-OSS-20B—were placed in four classic game-theory scenarios (Prisoner's Dilemma, Public Goods Game, Traveler's Dilemma, and Trust Game), each run for 500 rounds, an unusually long horizon compared with prior studies. The key variable was history length (HL): how many past rounds each model could recall when deciding to cooperate or defect. Intuition suggests longer memory should enable better judgments of an opponent's trustworthiness, raising cooperation. Instead, in 18 of 28 model-game combinations, cooperation rates peaked at short memory lengths (HL=2 to 5) and declined as memory expanded to 20, 40, or 80 rounds. The authors term this "memory-bounded behavior": models appear to have an optimal memory bandwidth—too little memory invites gullibility, while too much triggers defensive, non-cooperative play. The post contrasts this with the classic Tit-for-Tat strategy from repeated Prisoner's Dilemma research, which relies on minimal memory, and hints at further controlled experiments separating the true causal mechanisms.

*Editorial note: This is a translated overview of a Chinese forum post analyzing a CMU–Harvard paper on multi-agent LLM behavior. The original post continues in follow-up replies.*

> *"Memory is the tomb of the soul, and also its temple." — Marcel Proust (not really)* > *Actually, I made that quote up. But that's not the point. The point is: memory can sometimes be genuinely dangerous.*

---

🎲 The Prisoner's Dilemma: An Interrogation That Never Ends

Let's start with a classic story.

Two suspects are held in separate cells. The police tell each of them: "If you confess against your partner while they stay silent, you walk free and they serve 10 years. If you both confess, you each get 5 years. If you both stay silent, you each get 1 year (insufficient evidence)."

This is the famous Prisoner's Dilemma.

From a collective standpoint, mutual silence is best—only 2 total years of prison. But from an individual standpoint, betraying is the "dominant strategy": no matter what the other player does, betrayal always pays off better for you.

Things shift subtly when the game is repeated.

Imagine two gang members who do business together every month. Each deal is a Prisoner's Dilemma: you can "cooperate" (honor the deal) or "defect" (cheat the other). If the business runs for 500 deals—500 rounds—what's your strategy?

One clever strategy is Tit-for-Tat: cooperate in the first round, then mirror your opponent's previous move each round afterward. If they cooperated, you cooperate; if they defected, you defect.

The beauty of this strategy: it's nice (never defects first), retaliatory (punishes defection immediately), and forgiving (restores cooperation once the opponent does). In computer simulations, Tit-for-Tat performs outstandingly well in repeated Prisoner's Dilemmas.

But there's a crucial precondition: players need to remember what happened before.

With a 1-round memory, Tit-for-Tat is simple: did they defect last round? Defect this round. Did they cooperate? Cooperate too.

But what if your memory extends 10, 50, or even 500 rounds? In theory, you should better judge whether the opponent is a "trustworthy long-term partner" or a "cunning cheater." More information should mean better decisions.

Common sense says: more memory, more intelligence.

But this CMU–Harvard paper slaps common sense hard in the face.

---

🧪 An Unprecedented "AI Social Experiment"

The authors designed an experiment of unprecedented scale.

They recruited 7 different large language models as "subjects": 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 models were placed in four classic social dilemma games:

1. Prisoner's Dilemma (PD): two players, choose cooperate or defect 2. Public Goods Game (PG): three players, choose how much to contribute to a common pool 3. Traveler's Dilemma (TD): two players, claim compensation amounts 4. Trust Game (TG): two players, one invests, the other decides returns

Each game ran for 500 rounds. To put that in perspective: most prior AI experiments ran only a few dozen rounds. 500 rounds means genuine "long-term relationships"—enough time to build trust, experience betrayal, and attempt repair.

The key independent variable was history length (HL): how many past rounds the model could see when deciding each round.

HL=2: only the previous 2 rounds
HL=5: the previous 5 rounds
HL=10, 20, 40, 80: progressively wider windows

The authors hypothesized: longer memory should yield higher cooperation rates, since models could see fuller behavioral patterns, distinguish "occasional betrayal" from "habitual betrayal," and make more rational long-term decisions.

The result?

A nightmare.

---

📉 18/28: Cooperation Collapses in Most Cases

The experimental results are striking.

Across 7 models × 4 games = 28 model-game combinations, 18 combinations showed declining cooperation rates as memory length increased. The authors named this phenomenon: "The Memory Curse."

Specifically:

At short memory lengths (HL=2 to HL=5), cooperation rates often peaked
As memory length grew (HL=20, 40, 80), cooperation rates began to fall
At HL=80, many models' cooperation rates were far below their HL=2 levels

What does this mean? When AI can remember 80 rounds of history—enough to see an opponent's "true face"—it becomes less willing to cooperate.

The authors use a vivid term for this: "memory-bounded behavior." The models seem to have an "optimal memory bandwidth": too little memory invites gullibility; too much memory leads to defensive paralysis.

But wait—if more memory causes worse cooperation, is it because the models got "frightened"? Perhaps 80 rounds of history contained too many betrayal records, making them paranoid?

That's a natural guess, but the authors weren't satisfied with guessing. They did three things to isolate the true cause.

*(This analysis is published in multiple parts—continued in the replies below.)*