A tiny town where LLMs pretend to be people

What it does

This is the code behind the Generative Agents paper. It runs a browser-based simulation, Smallville, where LLM-backed agents wander a 2D map, maintain a memory stream, and generate believable behaviors. You pick a starter scenario, let the clock run in ten-second steps, and watch the agents generate behaviors that you can later replay or compress into a demo.

The interesting bit

The agents don’t just roleplay in a vacuum; they carry authored history files—semicolon-separated memory records injected at boot—that populate their memory stream. The whole system is essentially a Django frontend bolted to a Python backend that repeatedly asks an LLM what a pixel person should do next.

Key highlights

• Ships with two base scenarios: a 25-agent town or a focused three-agent setup featuring Isabella, Maria, and Klaus.
• Agent histories are loaded via CSV files that inject semicolon-separated memory records into each agent’s memory stream.
• Raw replays are explicitly labeled as debugging aids where every sprite looks identical; proper demos require a separate compression step.
• Expanding the town or renaming characters means editing the map in the Tiled editor and authoring new base simulation folders.

Caveats

• OpenAI API rate limits can hang the simulation, forcing a restart; the authors note that running many agents was “somewhat costly” even in early 2023.
• Firefox may produce frontend glitches; Chrome or Safari are the recommended browsers.
• The README reads like a quickstart manual, so the deeper mechanics of memory retrieval and behavior generation are left to the paper.

Verdict

Worth cloning if you’re studying LLM-based human behavior simulation or need a reproducible baseline for multi-agent experiments. Avoid if you want a polished engine or anything that runs cheaply without an OpenAI key.