Skip to content

Features

World Simulation Engine is built around four related design choices:

  • Time advances inside the simulated world, not just by chat-message count.
  • The world state is stored as typed graph data, not as one transcript or a set of disconnected tables.
  • Simulation is split across multiple focused LLM agents, each with a narrow job and its own model configuration.
  • Characters act from scoped perspectives, with private memories and beliefs instead of omniscient access to canonical state.

These features are implemented together. The scheduler asks who is due to act at the current simulation time. Each actor receives a perspective built from graph state, perception resolution, relationships, memory recall, intents, and private emotional constraints. The selected agents propose, validate, coordinate, narrate, and commit only structured changes. The graph then becomes the next source of truth.

flowchart LR
    Time["Simulation time"] --> Actors["Due actors"]
    Graph["Graph state"] --> Perspective["Actor perspective"]
    Memory["Private memory and beliefs"] --> Perspective
    Actors --> Agents["Focused LLM agents"]
    Perspective --> Agents
    Agents --> Commit["Validated state commit"]
    Commit --> Graph
    Commit --> Time

Read these sections for the details: