axiom-fabric

The governance ledger for agentic AI.

Agent memory systems answer "what do I remember that's relevant?" Axiom Fabric answers a different question: "what is true, how do I know, what breaks if it changes, and who authorized the change?"

It is a versioned, append-only truth ledger for the facts your agents and applications act on: every fact carries explicit provenance (derivation edges to the exact upstream fact-versions it was built from), an explicit trust weight (0–100, by layer policy), and a full version history. Updates append, retractions append tombstones — nothing is ever silently rewritten, so the history is the audit trail.

This package is the core: the SQLAlchemy data model (layers, layer_versions, facts, fact_versions, fact_version_edges), the repository functions all frontends share, Alembic migrations, the af command-line interface, and the af-mcp MCP server that lets agents (Claude Code, Claude Desktop, Gemini CLI, Codex CLI, …) read and write the ledger during execution.

pip install axiom-fabric          # `af` CLI + `af-mcp` server (MCP built in) — SQLite, zero infrastructure
pip install "axiom-fabric[llm]"   # + Anthropic / OpenAI extras

af init                           # apply migrations -> clean store (`--demo` seeds example layers)
af fact create --layer canonical --content '{"rule": "max_refund_usd", "value": 500}'
af fact list

To wire it into an MCP-capable agent (read tools always on; write tools only with --allow-writes):

af-mcp install --client claude --allow-writes --with-skill

Storage is SQLite by default (./af.db); set AF_DATABASE_URL to a postgresql+psycopg:// URL for production. The web dashboard lives in the separate axiom-fabric-dashboard package.

Full documentation, the design model, and the roadmap (cascade staleness, change-cost pricing, gated promotion of LLM output into truth) live in the project repository's README.md and build.md.