Technical Summary: The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents

Problem Statement

The integration of Large Language Models (LLMs) into multi-agent systems has introduced a central challenge: navigating social dilemmas (e.g., Prisoner's Dilemma, Trust Game) where individual short-term rationality conflicts with collective long-term welfare. Classical game theory (e.g., the Folk Theorem) suggests that expansive interaction histories and infinite horizons facilitate sustained cooperation. However, behavioral psychology suggests that excessively long memories can degrade cooperation in humans by overwhelming them with historical noise and fostering grudge-holding.

This paper investigates a critical gap: does expanding an LLM agent's memory (context window) build necessary trust, or does a perfect historical ledger inevitably doom the system to unforgiving retaliation? While prior work has evaluated LLMs in repeated games, these studies typically utilize very short horizons (≤10 rounds), potentially obscuring the effects of long interaction histories.

Methodology

The authors conducted a large-scale empirical study involving 7 representative LLMs (Gemma-3-12B, GPT-OSS-20B/120B, Llama-3.3-70B, Llama-4-Scout-17B, Mistral-7B, Qwen2.5-Coder-32B) across 4 classic social dilemma games (Prisoner's Dilemma, Traveler's Dilemma, Public Goods Game, Trust Game).

Experimental Design:

• Scale: Interactions were extended to 500 rounds, significantly longer than prior evaluations.
• Memory Manipulation: The accessible interaction-history length (HL) was varied across 9 settings: {0, 1, 2, 3, 5, 10, 20, 40, 80}.
• Reasoning: Agents were instructed to use Chain-of-Thought (CoT) reasoning to analyze game rules, history, and long-term objectives before acting.
• Data Volume: The study generated over 378,000 reasoning traces across 28 model-game settings (7 models × 4 games × 9 history lengths × 3 seeds).

Analytical Approaches:

1. Lexical Analysis: Automated classification of reasoning traces into "Forward-Looking" (proactive, long-term) and "History-Following" (reactive, risk-averse) categories.
2. Memory Sanitization: Replacing visible historical records with synthetic cooperative data while keeping prompt length fixed to isolate content effects from length effects.
3. Asymmetric Memory: Pairing "forgiving" agents (short memory, HL=2) with "grudge-holding" agents (long memory, HL=80) to observe systemic contagion.
4. Targeted Fine-Tuning: Training a LoRA adapter on a "Memory Cursed" model (Mistral-7B) using only traces enriched with forward-looking reasoning to test causal intervention.
5. Ablation Studies: Comparing agents with explicit CoT reasoning against those outputting only final decisions (No Reasoning).

Key Results

1. The "Memory Curse" Phenomenon

The study identifies a pervasive "memory curse" where expanding accessible history degrades cooperation in 18 out of 28 model-game settings.

• Behavioral Trajectory: Cooperation often peaks at short memory lengths (typically HL ≤ 5) and collapses as history expands. For example, in the Trust Game, Gemma-3-12B's cooperation rate dropped from 51.2% (HL=2) to 9.5% (HL=80).
• Variance: Longer memories amplify the "butterfly effect" of early noisy interactions, trapping agents in path-dependent retaliation cycles and increasing behavioral variance across seeds.

2. Cognitive Divergence: Immune vs. Cursed

Not all models succumb to the curse. The authors identified two distinct regimes:

• Memory Immune (10 settings): Models like Llama-3.3-70B and Qwen2.5-Coder-32B in specific games maintained ≥95% cooperation across all history lengths.
• Memory Cursed (18 settings): These models exhibited severe decay.
• Mechanism: Lexical analysis revealed that "Immune" models sustain a significantly higher ratio of forward-looking reasoning (e.g., "future," "long-term," "mutual benefit") even at HL=80, whereas "Cursed" models shift toward history-following and defensive reasoning (e.g., "risk," "betray," "retaliate").

3. Content vs. Length

• Sanitization: When 78 of 80 history rounds were replaced with synthetic cooperative records (keeping prompt length at 80), cooperation was substantially restored. This proves the trigger is memory content (accumulated negative evidence), not merely the context window length.
• Asymmetric Contagion: In mixed settings, a single "grudge-holder" (HL=80) drags down the group welfare, while a "forgiver" (HL=2) can maintain cooperative intent even when outnumbered by grudge-holders.

4. The Role of Reasoning

• Paradoxical Amplification: Explicit Chain-of-Thought (CoT) reasoning often amplifies the memory curse. In some cases (e.g., Llama-3.3-70B in the Trust Game), removing CoT increased cooperation from 6.9% to 100% at HL=80.
• Mechanism: CoT traces often devote space to enumerating past defections, making retaliatory decisions easier to justify and hardening the interaction into a negative pattern.

5. Intervention via Fine-Tuning

• Causal Proof: Fine-tuning a "cursed" Mistral-7B model exclusively on forward-looking reasoning traces (without explicit action labeling) successfully mitigated the decay.
• Zero-Shot Transfer: The fine-tuned model achieved near-100% cooperation at HL=80 across all four games, including those not seen during training, suggesting the intervention instilled a transferable reasoning tendency rather than memorizing specific game actions.

Significance and Claims

The paper claims to provide the first systematic sweep of interaction memory length in repeated LLM social dilemmas at this scale. Its primary contributions are:

1. Recasting Memory: It challenges the assumption that longer context windows naturally enhance strategic cooperation. Instead, raw memory expansion can become a liability, destabilizing trust by exposing agents to accumulated negative evidence.
2. Mechanism Identification: The "memory curse" is driven by a cognitive vulnerability to historical noise, where agents shift from forward-looking intent to defensive, history-following reasoning. This is paradoxically exacerbated by explicit deliberation (CoT).
3. Interventional Evidence: The study demonstrates that the curse is not an inevitable architectural limitation but a behavioral susceptibility that can be mitigated by explicitly anchoring an agent's reasoning toward long-term cooperation.
4. Future Directions: The authors suggest that for AI agents in long-horizon collaborative settings, equipping them with the cognitive capacity to "forgive" (via selective forgetting, summarization, or forward-looking reasoning priors) is as critical as expanding raw memory limits.

The paper concludes that while memory is essential for reciprocity, uncurated, verbatim historical recall in LLM agents can lead to algorithmic dysfunction, necessitating complementary mechanisms to filter historical noise and preserve the possibility of future trust.