Modern AI systems increasingly depend on persistent representations-world models, retrieved facts, and evolving knowledge structures-to support reasoning, planning, and action over time. Yet persistence does not guarantee that stored information remains justified when observations change, documents are updated, or task requirements shift. We argue that a central unresolved problem in AI memory is not merely storage or retrieval, but maintenance: determining which stored distinctions remain epistemically supportable under current evidence and task demands, and which of those supportable distinctions can still be materialized under finite compute, memory, latency, and energy budgets. These are distinct failure modes: epistemic loss calls for reobservation or revalidation, whereas resource-induced loss calls for compression, deferral, or reallocation. We advocate an explicit epistemic maintenance layer that continuously tracks both the supportable boundary and the resource-feasible maintained boundary of persistent memory. We introduce boundary-conditioned epistemic memory (BCEM) as one architectural realization of this principle, and describe memory homeostasis as the closed-loop control mechanism that keeps memory aligned with changing evidence, tasks, and resources. This perspective reframes AI memory from passive storage to active maintenance of what the system remains justified-and able-to know.