“Generative Agents: Interactive Simulacra of Human Behavior,” posted to arXiv on April 7, 2023 by Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein - a Stanford and Google team - populated a small sandbox town, nicknamed Smallville, with twenty-five characters whose behavior was driven entirely by a language model. The paper became famous for one emergent result: told only that a single agent wanted to host a Valentine’s Day party, the agents independently spread invitations, made new acquaintances, asked one another on dates, and showed up at the right time and place over two simulated days.
The contribution underneath the demo is an agent architecture for believable long-term behavior. Each agent keeps a memory stream of its experiences in natural language; a retrieval step surfaces the relevant memories given the current situation; a reflection step synthesizes those memories into higher-level conclusions; and a planning step turns conclusions into concrete actions. Ablating any of these components - observation, planning, or reflection - made the agents noticeably less coherent.
Generative agents mattered because they reframed an LLM not as a chatbot but as the cognition inside a persistent character that remembers, reflects, and acts over time. The work seeded a wave of multi-agent simulations, NPC and game research, and social-science experiments, and it is one of the most-cited demonstrations that simple language-model loops can produce surprisingly lifelike collective behavior.