Utopia: Open-Source Enterprise World Model for AI Agents with Bitemporal Graphs

RAG Finds Documents, But Cannot Preserve Historical Truth

Retrieval-Augmented Generation (RAG) excels at fetching current policy snippets, yet fails when an agent must answer questions about a past business state — such as which payment terms applied to a contract on March 15, 2024, after multiple amendments. Utopia's README illustrates this with a Master Services Agreement example: the system must return the effective terms on that date and cite the exact amendment document. If the knowledge layer only stores the latest value, a later audit of a February pricing decision would incorrectly use the current customer tier (e.g., Enterprise) instead of the tier at decision time (SMB).

Bitemporal Graph: Two Timelines for Every Fact

Utopia's core data structure is a Bitemporal Graph . Each fact carries two independent timelines:

Valid time — when the fact was true in the real world (e.g., contract effective March 1).

Transaction time — when the system recorded or corrected the fact (e.g., ERP synced on March 5; amount corrected in May).

Instead of overwriting, corrections end the old fact's validity interval and create a new version with provenance links. An agent can therefore query "What did the system believe the contract value was on March 10?" and receive a versioned answer with source references.

Figure: Utopia's bi-temporal graph and derivation model

Conflict Detection, Ontology, and Human Review Queue

Enterprise data arrives from CRM, ERP, contracts, emails — often contradictory. Utopia integrates:

Ontology — defines entity types, cardinality constraints, transitive/symmetric/inverse relations.

Entity Resolution — matches by name, aliases, then embedding similarity; unresolved cases go to a Review Queue for human judgment.

Conflict Handling — options include closing the old fact, keeping both versions, rejecting the new fact, or (for ontology violations) retracting facts, modifying rules, or accepting the conflict.

Forward-Chaining Derivation — rules compile to derived facts, but derivation is disabled by default because erroneous rules propagate errors. Derived facts are tagged with their proof; explicit facts always override derived ones.

This pipeline acknowledges that enterprise knowledge is messy and that automated resolution cannot replace governed human oversight.

MCP Integration: Reads Are Typed and Dated, Writes Are Proposals

Utopia exposes an MCP (Model Context Protocol) Server so clients like Claude Desktop or Cursor can query the knowledge base with authenticated tokens. The design principle: Reads are typed and dated, writes are proposals. Agents receive structured, time-stamped facts — not raw text chunks — and any mutation enters as a Proposal requiring human confirmation before commitment.

A Decision Ledger (append-only) records every confirmation, rejection, merge, or graph rebuild with actor, timestamp, and target object. History is never overwritten, enabling audit trails for questions like "Why did we believe this fact? Who approved the change? Was the state at decision time identical to today's view?"

"Enterprise World Model" Is a Direction, Not a Mature Category

The term borrows from RL's World Models (Ha & Schmidhuber, 2018), but Utopia does not simulate enterprise dynamics or predict future states. It builds a queryable, updatable, versioned, and provenance-tracked business state layer — customers, contracts, organizations, their relationships, and the full evolution of each fact.

Current maturity: v0.1 . The project warns that database schemas may change, migrations are forward-only (no rollback), and production use requires version pinning and backups. Roadmap items like Decision Intelligence, decision replay, Execution Gate, Business Rules, and Agent Memory over MCP remain under development.

From Static Context to Continuously Maintained Business State

Agent context has evolved from document retrieval to semantic layers, knowledge graphs, and now a unified substrate that combines ontology, temporal versioning, conflict resolution, permissions, and audit logs. The practical shift: context is no longer a static bundle of passages for a model's context window, but a living, governed business state. Whether Utopia can stay accurate, maintainable, and free of operational overhead beyond v0.1 is the key open question.