City-scale multi-agent simulation with deterministic replay, a layered memory stack, and 28 research-grounded personality archetypes. Open methodology, reproducible artifacts, commercial tiers.
$ pip install luvoire-engine
A three-layer stack — working, episodic, long-term — persisted across sessions, with forgetting dynamics.
Every tick is recorded to msgpack. Same seed, byte-identical. Citable and auditable.
28 research-grounded archetypes (CAT-28) plus criminology Tier 1/2. Each card carries its own bibliography.
Four engine guarantees, written as primitives rather than marketing. Each maps to a public artifact, benchmark, or citation.
Same seed, byte-identical msgpack. Verified at 10,000 agents on a 60×60 grid, weekly regression guarded at 15%.
Learn more →Working, episodic, and long-term memory modeled separately. Session-persistent with explicit forgetting dynamics.
Learn more →CAT-28 (22 single + 6 composite) plus criminology Tier 1/2 types. Each with DOI/ISBN/ISSN references in the methods appendix.
Learn more →CycloneDX SBOM, attribution, audit log, reproducible artifacts. Written to clear research review and compliance.
Learn more →A recorded run of 10,000 agents on a 60×60 grid, Gangnam 7 p.m. scenario. Every frame is deterministically reproducible from the seed below.
Citations, provenance, artifacts, and licensing written to hold up in peer review and in a commercial procurement pass.
CAT-28 rests on twelve peer-reviewed citations and six monographs. Every archetype card surfaces its DOI, ISBN, and ISSN inline.
Nemotron-Personas-Korea from NVIDIA, 2025, CC-BY-4.0. Grounded in KOSIS demographics. Revision SHA pinned per release.
Five public msgpack runs, SHA256 fixed, guarded against 15% regression by a weekly CI run. Replay on any machine, any OS.
MIT core with commercial tiers. Toss Payments for KR, Paddle globally. PIPA and GDPR-aligned privacy drafts available.
Full comparison table lives on the pricing page. These are the core numbers.
For coursework, replication studies, and first runs.
For the paper you're writing this semester.
For labs and small studios.
For R&D labs and policy teams.
% BibTeX @software{luvoire2026, title = {Luvoire}, author = {Choi, J.}, year = {2026}, doi = {10.5281/...} }How to cite →
# Unity adapter using Luvoire; var npc = new Agent( archetype: CAT28.Guardian, seed: 42 );Get a Team license →
# curl curl -X POST \ api.luvoire.com/v1/runs \ -d 'seed=42, agents=10000, scenario=crisis'Talk to sales →
Three idioms, one engine. Pick Python for notebooks, CLI for reproducible batches, REST for service integration.
A grid, an agent count, and a seed fully specify a run.
Each tick advances decision, memory, and movement layers together.
Msgpack output is byte-identical for equal seeds — safe to cite.
Any machine, any OS. Replay verifies its own SHA256 on load.
import luvoire sim = luvoire.City( grid=(60, 60), agents=10_000, seed=42, ) replay = sim.run(ticks=100) replay.save("my_run.msgpack") # deterministic: same seed → same bytes assert replay.sha256() == "a9f3e8b2…c02137d4"