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

Forum topic · 小凯 · 2026-05-11

Summary

A paper from CMU and Harvard researchers reveals a counterintuitive phenomenon called the Memory Curse: in repeated social dilemma games, large language model agents cooperate less as their memory of past interactions grows. The authors tested 7 LLMs (Llama-3.3-70B, Qwen2.5-Coder-32B, Gemma-3-12B, GPT-OSS-120B, Mistral-7B, Llama-4-Scout-17B, GPT-OSS-20B) across four classic games (Prisoner's Dilemma, Public Goods, Traveler's Dilemma, Trust Game) for 500 rounds each, varying the length of visible history (HL=2 to 80). In 18 of 28 model-game combinations, cooperation rates dropped as memory lengthened, often peaking at short memory windows like HL=2-5. Analyzing over 378,000 chain-of-thought reasoning traces, the study found that longer contexts reduce forward-looking vocabulary (about future, reciprocity, trust) while defensive language does not increase. The mechanism is thus not paranoia but a crowding-out of the agents' capacity to envision mutual benefit: extensive history drowns the imagination of future cooperation. This 'memory-bounded behavior' suggests LLM agents have an optimal memory bandwidth, challenging the assumption that more context always improves decision-making in multi-agent systems.

The Memory Curse: When AI Remembers More, It Trusts Less

> *"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 and your partner stays silent, you go free immediately and he serves 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, both staying silent is best — only 2 total years in prison. But from an individual standpoint, betraying is the "dominant strategy": no matter what the other does, betraying always pays off more.

The picture changes subtly when the game is repeated.

Imagine two gangsters who do business together every month. Each deal is a prisoner's dilemma: you can choose "cooperate" (honor the deal) or "defect" (cheat your partner). If the partnership runs for 500 deals — 500 rounds — what's your strategy?

A clever strategy is Tit-for-Tat: cooperate in the first round, then mirror whatever your opponent did last round. If they cooperate, you cooperate; if they defect, you defect back.

The beauty of this strategy is that 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 exceptionally well in repeated prisoner's dilemmas.

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

With a memory of only 1 round — seeing just the previous outcome — Tit-for-Tat is simple: did he defect last round? Defect this round. Did he cooperate? Cooperate too.

But what if your memory extends to 10, 50, or even 500 rounds? In theory, you should be better at judging whether your counterpart is a "trustworthy long-term partner" or a "cunning defector." More information should mean better decisions.

Common sense says: more memory, more intelligence.

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

---

🧪 An Unprecedented "AI Social Experiment"

The paper's 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. These 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 the common pool 3. Traveler's Dilemma (TD): two players, claim compensation amounts 4. Trust Game (TG): two players, one invests, the other returns

Each game ran for 500 rounds. What does 500 rounds mean? Most prior AI experiments only ran a few dozen rounds. 500 rounds means a genuine "long-term relationship" — enough time for models to build trust, experience betrayal, and attempt repair.

The key independent variable was history length (HL): how many previous rounds the model can see when making each decision.

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

The authors hypothesized: longer memory should mean higher cooperation rates, because models could see more complete patterns of behavior, distinguishing "accidental betrayal" from "habitual betrayal" and making more rational long-term decisions.

The result?

A nightmare.

---

📉 18/28: Cooperation Collapses in Most Cases

The experimental results are stark.

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 grew further (HL=20, 40, 80), cooperation rates began declining
At HL=80, many models' cooperation rates were far below their HL=2 levels

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

The authors describe this with a vivid term: "Memory-Bounded Behavior." The models seem to have an "optimal memory bandwidth": too little memory invites gullibility; too much memory breeds defensiveness.

But wait — if more memory leads to worse cooperation, is it because the models got "frightened"? Maybe 80 rounds of history contain too many betrayals, making the models paranoid?

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

---

🔍 Mechanism One: Not Paranoia — "Foresight" Is Dying

The first thing they did was analyze the models' "thought processes."

Each model generates Chain-of-Thought reasoning when making decisions. The authors collected over 378,000 reasoning traces, then used semantic analysis tools to classify the vocabulary used.

They distinguished two categories:

Forward-looking vocabulary: "future," "long-term," "mutual benefit," "reciprocity," "building trust"
Defensive/History-following vocabulary: "retaliate," "punish," "he betrayed last round," "caution"

The results were striking.

As memory length increased from HL=2 to HL=80, the share of forward-looking vocabulary declined across the entire sample. In "immune" models (the 10/28 combinations unaffected by the curse), the forward-looking vocabulary ratio at HL=80 stayed at 0.504; in "cursed" models, it plummeted to 0.340.

More critically, the absolute frequency of defensive vocabulary did not significantly increase. The models didn't become "more paranoid" — they simply stopped imagining the possibility of future cooperation.

The authors' own words:

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

This is a profound insight. The essence of the Memory Curse is not "remembering too many bad things" but "the noise of historical information drowning out the imagination of the future." When 80 rounds of dense history sit before a model, its cognitive resources are occupied by the past, leaving it unable to think about "what if we keep cooperating for the next 500 rounds."

A metaphor:

Imagine meeting a new friend at a dinner party. If you only consider your most recent interaction — "last time he borrowed my notes and returned them" — you can easily judge him as reliable and cooperate again. But if someone hands you an 80-page dossier detailing all his past relationships — including stealing a classmate's eraser in elementary school, skipping one group meeting in college, yet donating to charity for three straight years — what then?

You'd likely fall into "analysis paralysis": too much information, no clean "trust/distrust" signal. Worse, any negative detail gets amplified (what psychologists call "negativity bias") while positive details drown in the noise. In the end, you might choose "don't risk it" — that is, don't cooperate.

An LLM at HL=80 is a victim of exactly that "80-page dossier."

*(Analysis continues in the reply below)*

#PaperReading #PapersCool #DailyPaper #MultiAgent #MemoryCurse