Repository navigation
Live demo: https://sidmohan0.github.io/rumorwoods/
A faithful, fully in-browser reimplementation of Generative Agents: Interactive Simulacra of Human Behavior (Park et al., UIST '23). Twenty-five agents perceive, remember, reflect, plan, and gossip — all cognition driven by a real LLM, no server, no API keys.
What's in this checkpoint
- The real Ville map — all 19 sectors, 63 arenas, 222 objects, and per-tile collision converted from the official repository's Apache-2.0 tile-matrix data; visuals re-skinned with Kenney's CC0 Roguelike pack (the original's commercial art is not included)
- Two backends — WebLLM (WebGPU, in a Web Worker) or any local OpenAI-compatible server (llama.cpp, Ollama, LM Studio, vLLM) with connectivity check and model picker
- Character editor — edit any resident's traits, story, and home, or add/remove residents live; custom rosters persist and embed in saves
- Sessions — save and resume full simulations in the browser (IndexedDB), including memory streams with embeddings
- Benchmark harness — headless (
npx tsx tools/benchmark.mts) and in-browser (?bench=1), with one measured game day: 1.47M tokens, 5,790 calls, 1.16h wall-clock, $0 vs. the original's 12.7M tokens/game-day on GPT-3.5 and 'multiple days' of wall-clock - Efficiency work — packed LLM calls, JSON-schema-constrained outputs, single serialized queue
Paper-parity notes
Memory stream with importance scoring, retrieval by recency × importance × relevance, reflection at the paper's threshold, hierarchical planning with lazy decomposition, perception with attention bandwidth, memory-conditioned dialogue, post-conversation memos, new-day identity revision, interviews, and the Valentine's party information-diffusion scenario. See the README's gap analysis for intentional differences.