How Multi-Agent AI Simulations Keep a Shared World State

What actually holds a simulated world together when multiple AI agents act inside it? This episode opens the repos on three frameworks that answer the question completely differently. Concordia from Google DeepMind makes world state a first-class Game Master component — a plain Python dict whose contents are authored by the model at every step, with a SQLite/Postgres storage protocol that ships as an interface but not an implementation. Generative Agents goes the other way: a deterministic tile grid in CSV files, per-agent JSON scratch files, and a memory system with a 0.995 recency decay. CAMEL has no persistent world object at all — state lives in the conversation history and a Task object. Along the way: why agent memory and world state must be separate, how entity-aware event filtering produces information asymmetry, and the four-line docstring about a rent event that silently refunded every agent's spending — and why the fix was atomic wallet deltas, not a better prompt. Key sources: • Concordia repo (google-deepmind/concordia), commit 2d0c2a4, README + CHANGELOG v2.4.0 (2026-03-06) — https://github.com/google-deepmind/concordia • `concordia/components/game_master/world_state.py` — https://raw.githubusercontent.com/google-deepmind/concordia/main/concordia/components/game_master/world_state.py • `examples/concordia_island/sim/world_state_protocol.py` (2026) — https://raw.githubusercontent.com/google-deepmind/concordia/main/examples/concordia_island/sim/world_state_protocol.py Episode 5915 — myweirdprompts.com/n/5915 (permanent short link: myweirdprompts.com/761242)