# Generative Agents: Interactive Simulacra of Human Behavior
## One-Page Summary

**Authors:** Park, O'Brien, Cai, Morris, Liang, Bernstein (Stanford & Google)  
**Published:** UIST 2023 | **Pages:** 22

---

## What Is This Paper About?

Stanford and Google researchers created AI agents that simulate believable human behavior over time. Unlike chatbots that respond in the moment, these "generative agents" remember experiences, reflect on them, and plan ahead.

---

## The Breakthrough

A 25-agent virtual town (inspired by The Sims) where:
- Agents wake up, go to work, talk to each other
- They form relationships and spread information
- They coordinated a Valentine's Day party from a single seed suggestion

All behavior emerged naturally from the architecture—no manual scripting.

---

## How It Works

Three components power each agent:

| Component | Purpose | Example |
|-----------|---------|---------|
| **Memory Stream** | Records all experiences | "Had coffee with Maria" |
| **Reflection** | Synthesizes patterns | "I enjoy discussing research with Maria" |
| **Planning** | Generates future actions | "Meet Maria at cafe tomorrow" |

The key insight: retrieval combines **recency**, **importance**, and **relevance** to surface the right memories at the right time.

---

## Results That Matter

- **Evaluation:** 100 human judges ranked the full architecture most believable (significantly beating ablations)
- **Information spread:** Party invitation reached 52% of agents in 2 days
- **Relationships:** Network density jumped from 17% → 74%
- **Cost:** Thousands of dollars in API tokens for 2 days of simulation

---

## Limitations Found

1. Agents sometimes embellish memories (hallucinate details)
2. Dialogue can feel overly formal
3. Location confusion as agents learn more places
4. Norm violations (entering occupied bathrooms, closed stores)

---

## Why This Matters

This moves beyond "chatbot in the moment" to agents with persistent memory, personality, and temporal coherence. Applications range from game NPCs to social prototyping to rehearsal spaces for difficult conversations.

---

## Key Takeaway

> "With generative agents, it is sufficient to simply tell one agent that she wants to throw a party... they spread the word and then show up."

The architecture handles the rest—no scripting required.