How memory can affect collective and cooperative behaviors in an LLM-Based Social Particle Swarm

Researchers demonstrated that memory length in LLM-based multi-agent systems produces contradictory effects on cooperation depending on the model used: Gemini showed suppressed cooperation with longer memory, while Gemma exhibited enhanced cooperation. The findings suggest model-specific characteristics and alignment mechanisms fundamentally shape emergent social behaviors in AI agent systems.

This research reveals a critical discovery in generative agent-based modeling: the relationship between memory and cooperation is not universal but deeply dependent on how individual LLMs process and interpret information. The study extends the Social Particle Swarm model—a established framework for studying collective behavior through game-theoretic interactions—by replacing rule-based agents with LLM agents. The divergent results between Gemini and Gemma models expose how internal model characteristics, potentially including alignment mechanisms, directly influence macro-level emergent behaviors.

The sentiment analysis findings provide crucial micro-level insight: Gemini increasingly interprets longer memory as negative context, progressively undermining cooperation, while Gemma maintains a more neutral interpretation that facilitates cooperative clustering. This distinction matters because it suggests LLM alignment techniques may inadvertently shape social dynamics in unexpected ways. For AI developers and researchers, this indicates that optimizing individual model behavior doesn't guarantee predictable multi-agent outcomes—system-level effects emerge from model-specific cognitive patterns.

The partial alignment between Big Five personality traits and observed agent behaviors validates the simulation's realism while raising important questions for AI safety. As LLM agents become deployed in collaborative systems, understanding how their internal characteristics propagate through multi-agent environments becomes essential. The study challenges the assumption that longer memory universally improves decision-making and cooperation, suggesting instead that implementation details and training choices create fundamental constraints on emergent behavior patterns.

• →Memory length effects on LLM agent cooperation are model-dependent, not universal across different LLMs
• →Gemini exhibits cooperation suppression with increased memory, while Gemma shows the opposite trend
• →Sentiment analysis reveals different interpretations of memory context between models, explaining behavioral divergence
• →LLM alignment and internal characteristics shape emergent social behaviors at the multi-agent system level
• →Current research contradictions on memory and cooperation may stem from unreported model-specific differences