Generative agents
Memory streams, reflection, and LLM-mediated planning create believable individuals — but a natural-language memory loop for every resident cannot scale to a city. Enreef keeps rich memory for selected agents and salient moments.
Research
Simulation you can interrogate.
Enreef optimizes for scientific auditability and replay, not black-box prediction claims. The team treats open questions as benchmarks to run, not debates to win. Publications and reproducible benchmarks will appear here as they exist — none are published yet, and we won't fake them.
What the field has learned
Enreef builds on — and deliberately diverges from — the current generation of agent and world-simulation research:
Compiled policies
Recent work grounds hundreds of thousands of synthetic residents in census data by compiling LLM-derived decision signals into lookup policies offline. Enreef adopts this as a core principle: reason expensively offline, then reuse.
Event-driven state
Naïve movement logging explodes — 500k simulated agents at 1 Hz over 70 days produce trillions of location records. Enreef stores transitions, trajectories by reference, and aggregates instead.
Microsimulation discipline
Decades of activity-based travel and land-use modeling provide validated synthetic-population and choice-model methods. Enreef borrows the discipline and moves the hot loop to a compiled runtime.
Grounded worlds are validated
The question: how closely does this simulation reproduce observed reality?
- Agreement with empirical observation
- Calibration & holdout periods
- Error metrics & benchmark models
- Uncertainty reporting
- Historical replay
Synthetic worlds are coherent
The question: is the world internally consistent, persistent, and faithful to its authored rules?
- Rule adherence & invariant checks
- Agent continuity & consistency
- Relationship coherence
- Resource conservation where required
- Long-horizon stability
Open questions
Each of these becomes a benchmark issue with a reproducible answer:
- Q01How fast can a single-node Rust engine simulate a full synthetic San Francisco before distribution is needed?
- Q02How much route choice can precomputed travel-time matrices serve without destroying scenario sensitivity?
- Q03What minimum set of behavioral variables materially improves on classic microsimulation?
- Q04How much does LLM reasoning improve holdout backtests per dollar?
- Q05Which historical San Francisco events provide clean holdout tests for mobility, housing, and business models?
- Q06How should world branches represent large state deltas most efficiently?
- Q07Can city-specific data sources become generic connectors instead of consulting code?