Skip to main content

World Model

The World Model holds the agent's structured view of real-world entities — currently per-entity snapshots and forecasts. It complements the temporal world model (timeline of observations) by storing the current state in a typed form.

Storage​

TablePurpose
EntityStateLatest snapshot per entity: current_value, previous_value, change_pct, trend, state_metadata, last_updated
StatePredictionLLM forecasts: entity, prediction_text, horizon, confidence, model_used, created_at, expires_at

EntityState is updated by the world_model_job (hourly) which aggregates the latest rows in WorldTimeline per entity.

StatePrediction rows are created by the dream cycle (when generating forecasts) and the perception job (when predicting near-term moves).

Entity types​

Tracked entities include:

  • Crypto assets (BTC, ETH, etc. — via WorldTimeline price extraction)
  • System metrics (CPU, RAM, latency, error rate, active users)
  • User-mentioned entities from the Knowledge Graph

Trend computation​

change_pct = (current_value - previous_value) / previous_value * 100
trend = "up" if change_pct > +5%
        "down" if change_pct < -5%
        "flat" otherwise

Tunable thresholds in memory/temporal.py.

Browse​

Dashboard /world-model shows:

  • Per-entity table with current value, change %, trend
  • Per-entity timeline chart (from WorldTimeline)
  • Active predictions

See also​