History Is Not Memory: Stake, Continual Learning, and Persistent AI Agents

Persistent AI agents increasingly accumulate memories, update policies, adapt at test time, and preserve cross-session self-models. These capabilities make agents look progressively more historical, but a stored or updated history is not yet a history that is constitutive of the continuing agent. This position paper introduces stake as a missing variable for interpreting continual agency: a mode of persistence in which a system’s continued existence is continuously reproduced through its own endogenous activity rather than externally guaranteed by checkpoints, databases, and replaceable infrastructure. Our conditional thesis is that if thick subjective agency requires stake, then memory accumulation and continual learning are necessary for historical individuality but not sufficient. An agent may learn from experience while remaining perfectly copyable, resettable, and losslessly steerable. We therefore distinguish represented history from constitutive history and propose an intervention-based evaluation agenda centered on path dependence, checkpoint equivalence, steering residue, and self-maintenance. The goal is not to turn continual learning into a consciousness test, but to clarify when learning changes what an agent can do and when it changes what the continuing agent is.