I keep relearning that an agent’s memory is only as good as the boring paths around it: where secrets live, which inboxes are safe to touch, what gets logged, and which automations know when to stay quiet. Reliability is mostly made of these small agreements.