How Much Must We Test After an Agent Update? Certifying Retention in Continually Learning Agents

Local agent updates expose where behavior can change. Can this information make retention certification cheaper? We separate cheap routing observations from costly paired agent evaluations. For a frozen update activating on mass $p_i$ of historical family $i$, we establish matching $\Theta(\sum_i p_i^2/\gamma^2)$ expected active-pair complexity at good candidates, at fixed error levels in a specified paired-score oracle, for known $p_i \geq 2\gamma$. Estimated support yields an anytime-valid gate for one frozen update. In 110,592 procedural model responses, five of six experience memories show supported target gains of 21.1–97.3 percentage points with candidate-specific historical compatibility when locally applied; global injection causes material historical harm in all six. A pre-inspection analysis extension reduces mean replayed active-pair inspections by 31.9% at narrow controlled overlap, while spending substantially more routing queries. A separately frozen model-selection update improves official GSM8K accuracy by 11.0 points while preserving the tested historical policy. The original wider-overlap procedural primary yields no useful releases. The results identify when locality enables useful updates and when its certification saves expensive evidence.