The persistent per-run record in StateM functions as an external recovery anchor during process restarts, context refreshes, or model-side compactions. Rather than forcing the agent to infer its current position from a noisy terminal transcript, the agent queries StateM to retrieve the authoritative phase, past transitions, unresolved checks, and evidence files.

The scope of this recovery is specifically procedural rather than transactional:

• Recoverable scope: StateM restores its recorded control state, active obligations, and configured repair routines.
• Non-recoverable scope: StateM cannot reconstruct unpersisted working memory, retrieve hidden model activations, or execute ACID-style rollbacks of arbitrary external environment side effects.

Reliability therefore depends on positioning persistence hooks at phase boundaries and ensuring hooks are idempotent to tolerate re-execution.