Skip to main content
A verified incident exports an engram.memory-regression artifact. Version 2 declares semantic lifecycle and answer assertions plus a controlled perturbation matrix, without claiming access to hidden model reasoning.

Run with the Studio executor

Your project owns the executor boundary. The recommended LangGraph setup puts that module in engram.config.json, so Studio and CI import the same code:
For custom integrations, an executor can also be a legacy function:
Execute the artifact:
With project configuration, omit both paths and run the entire suite:
The command invokes the executor once per matrix variant and exits nonzero when any required selection, exclusion, context load, or answer assertion fails. Failed findings identify the variant and show expected and observed values.

Check a recorded observation

When another system already produced the matrix results, compare the observation array directly:

Run in CI

The repository includes a complete copyable workflow at examples/github-actions/memory-regression.yml.
Version 1 artifacts and their ID-based executor contract remain supported. A version 2 artifact proves that every supplied matrix variant passed observable semantic assertions. Production equivalence exists only when the executor calls the production-equivalent retrieval and generation stack.