Underlying paper - Generative Agents

Generative Agents introduces a novel architecture for creating believable software agents that simulate human behavior. These agents are powered by large language models extended with a memory stream that stores experiences in natural language, a reflection mechanism that synthesizes memories into higher-level insights, and a dynamic planning system that retrieves relevant memories to guide future actions. The agents inhabit an interactive sandbox environment inspired by The Sims, where 25 autonomous agents wake up, go to work, form opinions, initiate conversations, and coordinate social events—all driven by their internal memories and plans. End users can interact with agents using natural language. Evaluations demonstrate that the agents produce believable individual behaviors and emergent social dynamics, such as autonomously planning and executing a Valentine's Day party. The architecture's components—observation, planning, and reflection—are each shown to be critical for believable behavior. This work provides architectural and interaction patterns for enabling realistic human behavior simulations in applications like immersive environments, prototyping tools, and social rehearsal spaces.