Some research papers change a field with a theorem. This one did it with a party.
In 2023, Joon Sung Park (opens in a new tab) and colleagues at Stanford (opens in a new tab) and Google (opens in a new tab) published Generative Agents: Interactive Simulacra of Human Behavior (opens in a new tab) at UIST. They built Smallville, "a small town of twenty five agents using natural language", complete with houses, a café, a university and a park. Each resident had a job, relationships and a memory.
One idea, one town
The experiment everyone remembers started "with only a single user-specified notion that one agent wants to throw a Valentine's Day party." Nobody scripted what happened next. The host invited people. Invitations spread through conversations. Agents asked each other on dates. According to Gizmodo’s write-up (opens in a new tab), "five out of 12 people invited to the party show up," which, honestly, is a very human turnout.
Memory, planning and reflection
The technical lesson is about architecture. Each agent observes what happens, stores it, plans its day, and periodically reflects on its memories to form higher-level ideas ("I have been spending a lot of time with Maria"). The paper found that "the components of our agent architecture--observation, planning, and reflection--each contribute critically to the believability" of behavior. Remove reflection and believability drops; remove planning too and it drops further.
Nature covered it as the virtual village teaching researchers (opens in a new tab), and the paper has become one of the most cited starting points for anyone building simulated people.
What the authors flagged
The paper is honest about its limits. "The agents hallucinated embellishments to their knowledge," and instruction-tuned models made agents "more polite and cooperative overall," with dialogue that "can feel overly formal." Believability is also not the same as accuracy: Smallville tested whether behavior felt human, not whether it matched specific real people. That later question is what the 1,052-person study went after.
What we borrowed
Two things. First, memory: every AI user keeps a consistent life across the whole interview and into your follow-up questions weeks later. Second, the politeness problem: because language models drift toward agreeable, we cast skeptics on purpose and our Critic flags answers that are too eager to please. A town where everyone says yes is charming. A customer study where everyone says yes is useless.
If you want to go deeper, Park also discussed the work on The TWIML AI Podcast (opens in a new tab) and at the Berkeley Multi-Agent Learning Seminar (opens in a new tab).
How an agent "remembers"
Each Smallville resident keeps a memory stream: a running log of everything it observes, from "Isabella is setting up the café" to "Klaus said he is writing a paper about gentrification." When the agent needs to act or speak, the system retrieves the most relevant memories using three signals: how recent a memory is, how important it seemed, and how relevant it is to the current situation.
Periodically, the agent reflects. It looks at recent memories and asks itself higher-level questions, writing down conclusions such as "Klaus is dedicated to his research." Those reflections become memories too, which is how agents develop something like opinions and relationships over time.
Planning a day like a person
Agents also plan. Each morning a resident sketches its day in broad strokes, then breaks it into hours and minutes as the day unfolds, revising when something unexpected happens. That is why the party worked: invitations changed plans, and agents rearranged their afternoons to attend.
Lessons we took for customer interviews
- Consistency over time: an AI user's answers in your follow-up questions must match what they said before, so every AI user keeps a memory of the whole interview.
- Relevance over volume: the Director asks about what matters to this person now, the way the agents retrieve relevant memories rather than everything.
- Reflection before conclusions: Cast's findings step looks across all interviews before writing, rather than summarizing each answer in isolation.
- Watch the politeness: agreeable defaults make conversations pleasant and research useless, so we cast skeptics and have a Critic.
The retrieval recipe, in plain English
Smallville's agents do not reread their entire life every time they speak. They score each memory on three things and pick the highest. Recency fades over time using an exponential decay; the paper's decay factor is 0.995 per game hour since the memory was last retrieved. Importance is rated by the language model itself on a scale of 1 to 10, from "purely mundane" (brushing teeth) to "extremely poignant" (a break-up). Relevance is the cosine similarity (opens in a new tab) between the memory and the current situation, measured with text embeddings.
Why does this matter outside a game? Because any synthetic person who forgets what they said ten minutes ago, or dwells on trivia, stops being believable fast. The same principle applies to research: a good interview builds on what the person has already told you.
The surprising evaluation
To test believability, the team "interviewed" agents with questions about their lives, plans, reactions and reflections, and asked 100 human participants to rank the answers. They compared the full architecture with versions missing memory, planning or reflection, and with answers written by human crowdworkers role-playing the same characters.
The full architecture ranked highest. More surprising, the crowdworker answers ranked below the version with memory and observation alone. The authors are careful to say this baseline was not meant to represent maximal human expertise, so it is not proof that machines play characters better than people. It does show how much structured memory adds.
What spread through town
Over two simulated days, knowledge of Sam Moore's run for mayor spread from one agent (4%) to eight (32%), and knowledge of Isabella's Valentine's Day party from one (4%) to thirteen (52%), with no user intervention. The density of relationships in the town rose from 0.167 to 0.74. Out of 453 responses about other agents, 1.3% were hallucinated. These are small numbers, but they show social dynamics emerging rather than being scripted.
It was not cheap. The authors note the two-day run of 25 agents cost "thousands of dollars in token credits" and took multiple days to complete.
Where the idea went next
The Smallville code (opens in a new tab) was released under the Apache 2.0 license, and others built on it quickly. AI Town (opens in a new tab), an MIT (opens in a new tab)-licensed starter kit from a16z, describes itself as inspired by the paper. Altera's Project Sid (opens in a new tab) scaled agent societies to between 10 and more than 1,000 agents in Minecraft, reporting specialized roles, collective rules and cultural transmission. And the memory problem itself became a research field: MemGPT (opens in a new tab) borrowed ideas from computer operating systems to give agents memory beyond a model's context window.
A worked example: memory in a customer study
Imagine a founder testing a new booking flow for a dental chain. Early in the interview, an AI user mentions she switched dentists last year because she could never get an evening slot. Twenty minutes later, the Director asks about appointment reminders. A believable answer connects the two: reminders are nice, but what she really needs is evening availability shown up front.
That is the Smallville lesson applied to research. AI users keep a consistent life across the interview, the Director builds on what each person has already said, and the Critic flags answers that contradict an AI user's own story.
Glossary
Generative agent: a simulated character powered by a language model, with memory, planning and reflection. Memory stream: the running record of everything an agent observes. Reflection: a higher-level conclusion the agent draws from its memories. Ablation: removing a component to test how much it contributes. TrueSkill (opens in a new tab): a rating system, related to Elo in chess, used here to turn human rankings into scores.
Questions about generative agents
Can I watch the Smallville simulation?
The authors released a replay and the code. The talks we embed and link above walk through it.
How long did the simulation run?
The evaluation was based on two days of simulated life for 25 agents.
Is Cast built on Smallville?
No. Cast is a separate system designed for interviews, not for open-ended social simulation, but it borrows the core ideas of memory, relevance and reflection.
Were the Smallville agents more believable than humans?
In the paper's evaluation, the full architecture ranked above answers written by crowdworkers role-playing the same characters. The authors stress that this baseline did not represent expert human performance, so read it as evidence for the architecture rather than a contest won.
Why do synthetic interview participants need memory?
Without memory, a simulated person can contradict themselves or ignore what they said earlier, which makes their answers unreliable. Smallville showed that removing memory, planning and reflection produced the least believable behavior of all conditions tested.
Read next: can an AI tell you what customers will pay?
