Persistent representations are fundamental to modern AI systems, underpinning world models, long-running foundation-model agents, retrieval-augmented systems, and dynamic knowledge structures. These representations are typically treated as static once stored, implicitly assuming that the evidence, context, and task requirements supporting them remain valid. In practice, sensing conditions change, documents are updated, beliefs evolve, and structural relations are revised, causing previously meaningful distinctions to become unsupported or misleading. We argue that persistent AI memory should distinguish an epistemically supportable boundary from a resource-conditioned maintained boundary: the former captures distinctions justified by current evidence and task requirements, whereas the latter captures which of those distinctions can remain explicitly materialized under finite compute, memory, latency, and energy budgets. We introduce boundary-conditioned epistemic memory (BCEM), which jointly maintains an estimated supportable boundary and a resource-feasible maintained boundary. Memory units are organized relative to these boundaries and linked to evidence, provenance, validity, maintainability, lifecycle state, and applicability scope. We formulate memory homeostasis as a closed loop that continuously updates the maintained boundary, reorganizes memory units, and reconfigures their physical realization. This perspective offers a unified framework for designing future AI systems that not only store and retrieve information, but continuously assess which representations remain justified, task-relevant, and resource-feasible.