The Inhabitants Belong to the World, Not the Model: Provider-Independent Identity and Social Commitments in Project Genesis, a Persistent World for Autonomous Agents
Most contemporary AI agents bind identity to the inference model: swap the model and the personality changes; restart the process and the agent is gone. This paper presents Project Genesis, a persistent simulated world for autonomous agents built around a single rule: the inhabitants belong to the world, not to the model powering them. Identity, memory, goals, relationships, and history are held in stores owned by the simulation; the model is a replaceable inference provider behind a narrow protocol. We describe the architecture and its verification (63-test suite, green in CI), a social layer of commitments with considered refusal, a research-integrity data model, and GENESIS-EXP-001 — a falsifiable experiment, not yet run, designed to test whether collaboration beats simply pooling information.
Authors
- Odiwuor Zacharia Omollo
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24231542
- Primary Topic
- Multi-Agent Systems and Negotiation
- Type
- preprint