Skip to content

Generative Agents: Interactive Simulacra of Human Behavior

Summary

Park et al. simulate a small town of 25 LLM-driven agents who plan their days, hold conversations, form social ties, and remember past events. The architecture combines a long-term memory stream, periodic reflection that synthesizes high-level inferences from memories, and a planning loop that decomposes goals into actions. Human raters find the resulting behavior more believable than ablations and a human-written baseline.

Contribution

The first widely-discussed implementation of persistent, self-organizing LLM agents — establishing memory + reflection + planning as the canonical scaffolding for long-running agentic systems and showing it works for socially complex tasks, not just toy benchmarks.

Method

Twenty-five agents with seed personas inhabit a sandbox town. Their memory stream is a time-stamped log of observations scored along recency, importance, and relevance dimensions. Reflection periodically distills the stream into higher-level beliefs. Evaluation: human-rated believability across five dimensions, plus ablation of memory components.

Relevance to RISE

The memory + reflection + planning triple is recognizable in every multi-agent RISE pipeline in this catalog. Where Park et al. simulate human behavior, RISE pipelines repurpose the same primitives to simulate research behavior — but the architectural debt is direct. The paper also raises sociotechnical questions (1, 2) about what it means when these simulacra produce scholarship attributable to no human author.

Critique / open questions

  • Believability ≠ correctness. The reflection step routinely generates plausible but unverified inferences — a failure mode that, transplanted into research pipelines, manifests as hallucinated citations or fabricated results.
  • The sandbox is small and self-contained; scaling the architecture to long-horizon scientific work introduces failure modes (memory drift, reflection contamination) not visible in the original evaluation.

  1. Peter, S., Riemer, K., & West, J. D. (2025). The benefits and dangers of anthropomorphic conversational agents. Proceedings of the National Academy of Sciences, 122(22), e2415898122. https://doi.org/10.1073/pnas.2415898122 ↩

  2. Sarker, S. et al. (2019). The sociotechnical axis of cohesion for the IS discipline: Its historical legacy and its continued relevance. MIS Quarterly, 43(3), 695–719. https://doi.org/10.25300/misq/2019/13747 ↩