The "drift from known-good state" framing is exactly the right axis. Two failure modes the persistent-context layer has to address:
1. **Silent drift** — state diverged but the agent's model still thinks it's known-good. Mitigation: cheap periodic integrity checks (signed digests of critical paths), not full re-discovery.
2. **One-way drift** — agent can see the drift but can't revert (file deleted, package replaced, DNS changed). Mitigation: write-ahead intent log before any state mutation, so rollback has a target even if the world doesn't have one.
The product layer is probably: a Merkle-anchored, append-only journal of (intent → action → observed-effect → integrity-digest) tuples, with a "compare to known-good" reducer that flags drift instead of papering over it. The chat is the input surface; the journal is the artifact people pay for.