generative-agents-smallville game sourced
Generative Agents (Smallville)
A research sandbox in which 25 language-model agents in a Sims-like town keep natural-language memory streams, reflect, and plan their days, producing unscripted information spread and social coordination.
core loop
agents perceive nearby events and objects -> observations enter the memory stream -> retrieve memories by recency, importance, and relevance -> plan the day and react -> converse and act -> reflections summarize memories -> plans and relationships change
design notes
Park and colleagues replace authored behavior trees with a loop of language-model calls on gpt-3.5-turbo. Every observation is stored as text with timestamps; retrieval scores memories by recency (decay 0.995 per game hour), importance (rated 1-10 by the model when stored), and embedding relevance, each normalized and weighted equally. When the summed importance of recent events passes 150 (two or three times a day) the agent writes reflections, and plans are decomposed from a day outline into hour and 5-15 minute chunks. Seeding one agent with the intent to host a Valentine's Day party led to 13 of 25 agents hearing about it and five attending over two game days, at a cost of thousands of dollars in API credits. The open-source release steps the town 10 game seconds per step.
relations
8 stored edgesDaily schedules
Each morning the model drafts a day plan of about five to eight broad chunks, decomposed into hour-long and then 5-15 minute actions that agents revise when they react to events.
Gossip propagation
Information spreads only through agent conversations; over two game days awareness of a mayoral candidacy rose from 1 to 8 agents and of a party from 1 to 13, with 6 of 453 awareness answers hallucinated.
Memories and thoughts
Each agent's memory stream stores observations, plans, and reflections as text; retrieval ranks them by recency (0.995 decay per game hour), model-rated importance (1-10), and embedding relevance, and reflections fire when summed importance exceeds 150.
Agent inspection
Researchers interview agents and read their memory streams to evaluate believability.
Generative content on demand
Plans, dialogue, importance ratings, and reflections are all generated by gpt-3.5-turbo calls at runtime.
Natural-language interaction
Users address agents in natural language as a persona or as the agent's inner voice, and can rewrite the status of objects in natural language.
Relationship values
Agents form acquaintances through conversation; network density rose from 0.167 to 0.74 over two game days.
backlinks
computed from other recordsclaims
1 with sourcesIn the two-day end-to-end run, relationship network density rose from 0.167 to 0.74 and five agents attended the Valentine's Day party.