The History of AI

Generative agents

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Agents

Lineage
Acting in the world
Kind
Publication

Who

Stanford University, Google Research

What happened

Joon Sung Park and colleagues at Stanford and Google create a small virtual town with twenty-five agents driven by a language model, which remember what they experience, reflect on it, and plan their day in a believable way.

Why it mattered

Simulated agents with memory, reflection, planning, and a social life, built on a language model.

The full story

Simulating human behavior believably was a long-standing ambition of video games and the social sciences. Park and colleagues tested whether a language model could do it with an architecture that gave it memory and continuity.

Each agent records its experiences in natural language. Over time, the system distills those memories into higher-level reflections, such as what the agent thinks of another person, and retrieves the relevant memories and reflections to plan what to do. The twenty-five agents lived in an environment inspired by the video game The Sims, and users could talk to them in natural language.

The most cited example began with a single idea: that one agent wanted to throw a Valentine's Day party. Over the next two days the agents spread invitations, made new acquaintances, asked each other out on dates, and showed up together at the party at the agreed time, without anyone having programmed it.

The experiments showed that observation, planning, and reflection each contributed to the believability of the behavior. The paper was presented at the UIST 2023 conference and showed that a language model with memory could sustain coherent behavior over days.

Sources

Original source

How to cite

“Generative agents.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/generative-agents

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline