Stanford's Generative Agents simulate social behavior in a virtual town, with a three-layer memory system—perception stream, memory retrieval, and reflection layer—at the core of its architecture.