The Memory Curse: When AI Agents Remember More, They Trust Less

> *"Memory is the tomb of the soul, and also its temple." — Marcel Proust (not really)* > *Actually, the author 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

Two suspects sit in separate cells. Police tell each one: "If you confess and your partner stays silent, you go free and he serves 10 years. If you both confess, you each get 5 years. If you both stay silent, you each get 1 year on insufficient evidence."

This is the famous Prisoner's Dilemma.

Collectively, mutual silence is best—only 2 total years served. Individually, however, betrayal is the dominant strategy: no matter what the other player does, betraying pays more.

Things change subtly when the game is repeated.

Imagine two gangsters doing a monthly deal. Each deal is a Prisoner's Dilemma: you can cooperate (honor the deal) or defect (cheat your partner). If the deal repeats 500 times, what's your strategy?

One clever strategy is Tit-for-Tat: cooperate in round one, then mirror your opponent's previous move. If they cooperate, you cooperate; if they defect, you defect back.

Its beauty: it's nice (never defects first), provocable (retaliates immediately), and forgiving (restores cooperation once the opponent does). In computer simulations, Tit-for-Tat performs remarkably well in repeated Prisoner's Dilemmas.

But there's a critical premise: players must remember what happened before.

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

But what if 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 cheat." More information should mean better decisions.

Common sense says: more memory, more intelligence.

But a paper from CMU and Harvard gives common sense a brutal reality check.

An Unprecedented "AI Social Experiment"

The authors 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, cooperate or defect 2. Public Goods Game (PG): three players, 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 returns

Each game ran for 500 rounds. In AI research, most prior experiments ran only a few dozen rounds. 500 rounds means a genuine "long-term relationship"—enough time to build trust, experience betrayal, and attempt repair.

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

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

The authors hypothesized: longer memory should yield higher cooperation rates, since models could see fuller behavioral patterns and distinguish "occasional betrayal" from "habitual betrayal."

The result?

A nightmare.

18/28: Cooperation Collapses in Most Cases

The findings 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: "The Memory Curse."

Specifically:

• Cooperation rates often peaked at short memory (HL=2 to HL=5)
• As memory extended further (HL=20, 40, 80), cooperation began to fall
• At HL=80, many models cooperated far less than at HL=2

This means that when an AI can remember 80 rounds of history—enough to see an opponent's "true face"—it becomes less willing to cooperate.

The authors describe the pattern as "memory-bounded behavior": models seem to have an optimal memory bandwidth. Too little memory invites gullibility; too much triggers defensiveness.

But wait—if more memory causes worse cooperation, maybe the models got "frightened" by 80 rounds full of betrayal records, becoming 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 Dying

The first was analyzing the models' "thought processes."

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

They distinguished two categories:

• Forward-looking vocabulary: "future," "long-term," "mutual benefit," "reciprocity," "building trust"
• History-following/defensive vocabulary: "retaliation," "punishment," "they defected last time," "caution"

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

Crucially, the absolute frequency of defensive vocabulary did not significantly increase. The models did not 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 imagination about the future." When 80 rounds of dense history sit in front of a model, its cognitive resources are consumed by the past, leaving no capacity to think "what happens 500 rounds from now if we keep cooperating?"

A metaphor:

Imagine meeting someone new at a dinner party. If you only consider the most recent interaction—"last time he returned my borrowed notes"—it's easy to judge him trustworthy and cooperate again. But if someone hands you an 80-page dossier of all his past relationship details—including stealing a classmate's eraser in elementary school, skipping one group meeting in college, yet donating to charity for three straight working years—what happens?

You likely fall into "analysis paralysis": too much information, no simple trust/distrust signal. Worse, any negative detail gets amplified (psychologists call it "negativity bias") while positives drown in noise. Ultimately, you may choose "don't risk it"—that is, don't cooperate.

An AI model at HL=80 is a victim of that 80-page dossier.

*(Interpretation continued in the reply thread)*