Refresh by Decision-Impact, Not Uncertainty: Governing a Stale User Model in Lifelong Planning Agents

A lifelong agent should remember its user across sessions, sparing them from restating everything and keeping their plan and progress up to date. Academic planning and advising requires such a feature. A persistent user model goes stale as students drift, through changes in personal preferences and new academic progress. Refreshing that model by re-asking the user costs the user's attention, so a lifelong agent must \textbf{govern its memory}: decide when the stored model is too stale to act on, and what to refresh. The default critetion is \emph{uncertainty}, which refreshes what the agent is least sure of; we study a sharper one, \emph{decision-impact}: refresh a dimension only when a plausible new value would change the agent's certified decision. Our central point is that uncertainty is not commensurable across a real agent's heterogeneous memory channels, whereas decision-impact is measured based on change to the committed plan. We instantiate this as a long-term planning workspace that persists soft constraints, plan, and progress, over a verifiable neuro-symbolic planner on real institutional requirements, and evaluate it with simulated students who drift mid-degree. On this cross-channel problem, decision-impact gating reaches near- alignment at about 5 questions (regret 4.4±24.9 over 60 trajectories) while the uncertainty baseline does no better than never refreshing. We also find that refresh \emph{relevance}, instead of refreshing rate, governs alignment. However, the boundary is symmetric: decision-impact misses a drift that escalates beyond what the agent expects, which is exactly the case the uncertainty baseline catches, so the two are complementary. For academic planning, the whole loop runs over a verified planner, so each change comes with a checkable certificate: a proof that the new plan is valid, or that the goal is now impossible. In short, we contribute a memory-governance criterion (decision-impact over uncertainty), the commensurability argument behind it, and a lifelong academic planning agent that realizes it.