What Survives a Model Change? Memory, Identity, and Succession in Personal AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

What warrants treating a replacement model as the continuation of a personal AI? Persistent memory offers anapparent answer, yet copying the same history into several systems exposes its insufficiency. I distinguishnumerical identity, recognizable character, and the legitimate continuation of delegated assistance. Throughtwo thought experiments, I argue that informational continuity neither determines a successor's authority norenlarges the scope of an inherited instruction. Two copies may possess the same past while occupying differentroles; two authorized workers may each perform an individually permitted action while jointly exceeding amandate that allowed only one execution. I develop an account of authorized succession in which the successorcan interpret the inherited record, act under a continuing mandate, and remain answerable for the handover.The argument draws on ZAI, my personal AI project, and engages philosophical and technical accounts ofcontinuity. It allows distributed assistance and advance delegation while preserving a distinction betweensuccessful succession and the survival of a conscious subject. What ought to endure through replacement is anintelligible relation to the user's work, including the user's authority to revise earlier commitments.