Authors: Joon Sung Park, Joseph C. OBrien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein (Stanford University & Google Research) · arXiv: 2304.03442
Designed a three-part cognitive architecture combining an append-only Memory Stream (scored by Recency, Importance, and Relevance), recursive Reflection trees, and hierarchical Top-Down Planning.
Key Quantitative & Theoretical Takeaway: Ablating any of Observation retrieval, Reflection synthesis, or Hierarchical Planning significantly degrades believable emergent social coordination in multi-agent environments.
Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication. We introduce generative agents: computational software agents that simulate believable human behavior.
Each memory object m is scored as a weighted combination of exponential recency decay (0.995^Delta_t), LLM-evaluated intrinsic importance (1-10), and cosine similarity between query and memory embeddings, normalized to [0, 1].
Replace arxiv.org with nujan.app on any arXiv URL (for example, https://nujan.app/abs/2006.11239) to open the PDF and AI research partner side-by-side.