Generative Agents: Interactive Simulacra of Human Behavior

The paper introduces generative agents, computational software agents that use a large language model to simulate believable human behavior. The core method is an agent architecture that records the agent's full experience as natural-language memories, periodically synthesizes those memories into higher-level reflections, and dynamically retrieves relevant memories to plan actions and respond to the environment. The authors populate an interactive sandbox environment inspired by The Sims with twenty-five agents that users can interact with in natural language.

In evaluation, the agents produced believable individual routines and emergent social behaviors: from a single seeded idea that one agent wanted to throw a Valentine's Day party, the agents autonomously spread invitations, formed new acquaintances, asked each other on dates, and coordinated to attend together. Ablation studies showed that the observation, planning, and reflection components were each critical to believability. By fusing LLMs with interactive agents, the work introduced architectural and interaction patterns for believable simulations of human behavior.