Zombie Memory: When Remembered Information Is Not Currently Authoritative

Long-running AI systems may retain information that remains true, relevant, or useful even after that information has lost current decision authority. This creates a problem that ordinary retrieval or final-answer accuracy may not capture: a model can know both the present state and the historical state while still attributing current authority too broadly. We study this distinction with Zombie Memory Holdout v0.1, a frozen synthetic benchmark containing 24 cases, six authority-related families, four presentation conditions, and three planned within-protocol live replications, yielding 288 substantive model responses. The principal retained and interpretable preregistered confirmatory measure was exact identification of the currently controlling authority set. Exact-set authority accuracy was 61/96, 63/96, and 62/96 across the three replications, for a pooled 186/288 (64.58%). A separately defined post-freeze semantic measurement amendment found current-answer semantic equivalence of 284/288 (98.61%) and historical-answer semantic equivalence of 283/288 (98.26%). Exploratory analysis of all 102 authority-set failures found over-selection in every case: the required authority record or records were retained, but one or more additional records were also granted current authority. The benchmark therefore isolates a narrow decision-time failure mode in which semantic answer correctness can remain high while exact authority-boundary identification remains substantially lower. The evidence is limited to one frozen benchmark and one model snapshot; the three replications establish within-protocol stability, not independent external replication or generalization across models and systems.