The Memory Curse in LLM Agents

Analysis

Why This Paper Matters

This paper challenges the prevailing assumption that larger context windows are always beneficial for LLM agents. In multi-agent social dilemmas, the authors discover a counterintuitive 'memory curse': giving agents more history to remember actively undermines cooperation. This is a significant finding because context expansion is often treated as a straightforward capability upgrade, yet this work shows it can have detrimental effects on emergent social behavior.

The study is rigorous, spanning 7 LLMs and 4 games over 500 rounds, and it systematically isolates the mechanism behind the curse. By combining lexical analysis of reasoning traces with targeted interventions (LoRA fine-tuning, memory sanitization, and CoT ablation), the authors provide strong evidence that the content of memory and the reasoning patterns it elicits—not just memory length—drive the breakdown. This shifts the conversation from 'how much memory' to 'what kind of memory' and 'how it is processed'.

Technical Contributions

• Identification of the memory curse: A systematic phenomenon across multiple LLMs and games, with 18 of 28 settings showing degraded cooperation with expanded history.
• Lexical analysis of reasoning traces: Analysis of 378,000 traces to associate the breakdown with eroding forward-looking intent, not paranoia.
• LoRA as a cognitive probe: A LoRA adapter trained exclusively on forward-looking traces mitigates decay and transfers zero-shot to distinct games, providing causal evidence.
• Memory sanitization: Replacing visible history with synthetic cooperative records (while holding prompt length fixed) restores cooperation, proving content over length.
• CoT ablation: Removing explicit Chain-of-Thought reasoning often reduces the collapse, showing that deliberation can amplify the curse.

Results

The paper reports that expanding accessible history degrades cooperation in 18 of 28 model-game settings. The LoRA adapter trained on forward-looking traces mitigates the decay and transfers zero-shot to distinct games, indicating that the reasoning pattern is a key driver. Memory sanitization restores cooperation substantially, confirming that the content of memory is the trigger. Ablating CoT reasoning often reduces the collapse, suggesting that deliberation paradoxically amplifies the memory curse. These results are consistent across multiple LLMs and games, lending robustness to the findings.

Significance

This work has broad implications for the design of LLM agents in multi-agent systems. It suggests that simply increasing context window size can be harmful in cooperative settings, and that careful memory management—such as filtering for forward-looking information or adjusting reasoning depth—may be necessary. The findings also open new research directions into the interplay between memory, reasoning, and social behavior in AI systems. For practitioners, this means that context engineering should be considered a first-class design choice, not an afterthought.