Generative Agents: Interactive Simulacra of Human Behavior

Joon Sung Park, Michael Bernstein, Percy Liang and colleagues at Stanford and Google built Smallville, a small town in a sprite-based sandbox populated by 25 agents, each described by a paragraph of identity and relationships, all driven by gpt-3.5-turbo. Users can watch them, talk to them, or change their environment. The paper, presented at UIST 2023, is about the architecture that keeps those agents coherent over days, not just over a single reply.

The center is the memory stream: every observation an agent makes ("Isabella is setting out pastries", "the desk is unoccupied") is stored as a natural-language record with a timestamp. To decide what to put in the prompt, the agent retrieves memories by scoring each one on three things: recency, an exponential decay of 0.995 per game hour since the memory was last used; importance, a 1 to 10 poignancy rating the model assigns when the memory is created (brushing teeth is a 1, a breakup a 10); and relevance, the cosine similarity of the memory's embedding to the current query. Each component is scaled to the range 0 to 1 across candidates and the three are added with equal weight.

Two further mechanisms make the agents seem to think. Reflection runs when the summed importance of recent events passes 150, about two or three times a day: the model reads the 100 most recent memories, asks itself the three most salient questions it could answer, retrieves evidence for each, and writes five insights, with citations to the memories behind them, back into the stream. Reflections can build on other reflections, forming trees. Planning writes a day's agenda in five to eight broad chunks, then decomposes it into hour-long and then 5 to 15 minute actions, also stored in the stream, and re-plans when something happens.

In an evaluation, 100 people ranked agents' answers to interview questions for believability; the full architecture won, and removing reflection, then planning, then memory each made agents less believable. In the two-day simulation, news of a mayoral candidacy and a Valentine's Day party spread through conversations, relationships formed, and five agents turned up at the party.

You are unlikely to build a simulated town for a customer. You will build assistants and agents that remember interactions, and this paper's three-part retrieval score and its reflection step are the patterns most memory systems still use.

Retrieval is more than similarity. Recency and importance matter for memory in a way they do not for document search. A support agent should weigh last week's escalation above a routine ticket from last year, even when both match the query.

The weights decide behavior. Equal weights after min-max scaling can let a dramatic but irrelevant memory crowd out the relevant one. Our simulation below shows how easily it happens. Tune the weights on the customer's own questions.

Reflection is summarization with provenance. Writing conclusions back into memory, with references to the observations behind them, is a good way to keep long histories usable. Keep the citations so conclusions can be checked and corrected.

Cost scales with every agent and every step. Twenty-five agents for two game days cost thousands of dollars in tokens and took days to run. Budget memory operations per user before promising "remembers everything".