Utopia: Open-Source Enterprise World Model Adds Bitemporal Knowledge for AI Agents

RAG Finds Documents, But Historical Facts Need More Than the Latest Version

RAG retrieves relevant text for questions like "What is the refund policy?" but fails for temporal queries: which payment terms applied to a contract on March 15, 2024, after multiple amendments? Utopia's README illustrates a Master Services Agreement where the system must return the effective terms on a specific date and cite the source amendment. If the knowledge layer only stores the latest value, a customer upgraded from SMB to Enterprise in April will appear as Enterprise for all past queries, distorting retrospectives on February pricing decisions. Agents participating in contracts, risk, and operations need to know what facts the system believed at a given time.

Bitemporal Graph: Two Timelines for Every Fact

Utopia's core structure is a Bitemporal Graph . Each fact has two timelines: valid time (when the fact was true in the real world) and transaction time (when the system recorded or modified it). Example: a contract effective March 1, but ERP synced March 5; an amount error discovered in April, corrected in May. The system retains both the original erroneous record and the correction, ending the old fact's validity interval and linking the new version. This preserves the evolution of business state, enabling queries like "What amount did the system show on March 10?"

Conflict Resolution and Human Review Built Into the Pipeline

Enterprise data arrives from CRM, ERP, contracts, emails — often conflicting. Utopia chains Ontology, Entity Resolution, Conflict Detection, Reasoning, and a Review Queue . Documents are parsed per ontology; entities are matched by name, alias, then Embedding Similarity ; unresolved cases go to an LLM judge. Low-confidence extractions, duplicate suspects, and cardinality violations enter the Review Queue for human action. Conflicts are not auto-resolved by "latest wins"; options include closing the old fact, keeping both versions, rejecting the new fact, or, for ontology constraint violations, retracting facts, modifying rules, or accepting the conflict. Derived facts via forward-chaining (transitive, symmetric, inverse relations) are opt-in, marked with provenance, and explicit facts always override derived ones . This prevents batch errors from bad rules and keeps the knowledge base a versioned, evidenced business state rather than a static snapshot.

MCP Integration: Agents Read Typed, Dated State; Write Proposals

Utopia exposes an MCP Server so clients like Claude Desktop or Cursor can query the knowledge base with permissioned tokens. The design principle: "Reads are typed and dated, writes are proposals." Agents receive organized business state — entity identity, fact type, validity interval, source, replacement links — not raw text chunks. Writes enter as Proposals requiring human confirmation. An append-only Decision Ledger records every confirmation, rejection, merge, split, or graph rebuild with actor, timestamp, and target object; history is never overwritten by current state changes. This audit trail becomes essential when agents drive procurement, risk, or contract workflows and teams ask why a fact was believed, its source, who approved a change, and whether today's state matches the decision-time version.

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

The term evokes RL world models (Ha & Schmidhuber, 2018) that learn environment dynamics for planning. Utopia does not simulate enterprise operations or predict state evolution; it builds a queryable, updatable, time-versioned, provenance-tracked business state layer (customers, contracts, organizations, relationships, fact validity, system knowledge times, modification history). Roadmap items like Decision Intelligence, decision replay, Execution Gate, Business Rules, and Agent Memory over MCP remain in development. The project is at v0.1 : database schema may change, migrations are forward-only with no rollback, and production use requires version pinning and backups. The real shift: enterprise context for agents is moving from a document set for model reading to a continuously maintained, versioned business state with conflict handling, human review, and auditability — whether it earns the "World Model" label is secondary.