What makes agent memory actually useful: good retention, reliable recall, clear scope, and enough visibility to trust the system in real workflows.
