Cruxible: An Open-Source Governed Truth Layer for AI Agents

Cruxible is an open-source project built to address the unreliability of text-based memory systems—wikis, markdown files, vector stores—used by LLM agents. Its creator found that agents couldn't consistently reconstruct evidence or semantic relationships across sessions, forcing constant re-verification of already-stored knowledge.

The project provides a governed state layer that human and agent teams operate together on top of an explicit, typed ontology. Anything that can be derived deterministically is built that way, while judgment-based changes go through strict review queues with full attribution receipts before being minted into state.

Reads happen outside the LLM through a query engine or predefined named queries, eliminating the need to re-reference original sources (no RAG) once truth has been established. Because custom logic is pinned, snapshots remain fully reproducible. One example use case models a company's IT infrastructure deterministically, allowing a workflow to automatically infer which assets and teams a vulnerability might affect downstream.

Built in Python with SQLite under the hood and licensed under Apache-2.0, Cruxible runs locally through an isolated daemon and includes a separate read-only inspection UI. The author argues that solid internal truth models will become critical organizational assets in the AI era and is seeking community feedback and contributions.