Agentic AI governance in operation

Watch an agent cross the line. Then watch Mneme stop it.

Two runnable proofs show consequential architectural failures before they land. A third page defines the multi-agent enforcement contract honestly: what works now, what does not, and what proof is still required.

Runnable proofs

The change looks useful. The decision says otherwise.

Flagship 01 Runnable locally · Deny → pass

A second database sneaks in.

JSON becomes Postgres. Mneme stops the migration, then checks the safer retry.

If it lands One product. Two sources of truth.

Open the six-beat proof →

Flagship 02 Runnable locally · Deny → pass

The rulebook starts guessing.

A vector-search change lets agents find different decisions for the same task. Mneme blocks the fork.

If it lands Same request. Different rules. Conflicting code.

Open the six-beat proof →

Infrastructure contract Forward-looking · Target capability

Many agents. One rulebook.

The target: every agent gets the same decision, meets the same gate, and leaves a shared trace.

Inspect the target contract →

Shared policy version Actor + verdict identity Persistent correction trace

Try it yourself

Run a governed check locally.

Open source. Deterministic. No model call required.

pipx install "mneme-hq>=0.5.1"

Supporting examples

Same system. Smaller, useful checks.

PassKeep JSON. Ship the change.

Guidance steers the first attempt.

WarnNew package? Make it a decision.

Flag the dependency before it drifts in.

DenyThe code works. The boundary does not.

Stop direct database access in a service.

InfrastructureTurn ADRs into checks.

Compile the rules you already maintain.

Integration targetPut the gate before the tool call.

A real check; future SDK wiring.

An engineer reviewing architecture and product flows drawn on a whiteboard — Architecture begins as a plan people can inspect. Agentic governance keeps later changes tied to it.