Retrieval-Augmented Generation (RAG) treats memory as a retrieval problem: store content, embed it, fetch what is relevant. This suffices for static knowledge bases but breaks down for agents that learn over time, where stored information contradicts itself, beliefs evolve, and corrections must propagate through dependent reasoning. Current architectures commit a category error, applying uniform persistence semantics to information with fundamentally different epistemic status. We introduce Layered Epistemic Agent Protocol (LeAP), a framework that treats agent memory as an epistemology problem. LeAP formalizes stratified epistemic types with distinct evidence requirements: observations require no justification, claims require single sources, facts require corroboration, and beliefs require derivation chains. It defines warrant functions, coherence invariants, and a revision protocol satisfying AGM belief revision postulates. Coherence is decidable for finite stores. We present CITE (Context In Tiered Epistemology), a four-layer architecture implementing LeAP. Its layers (Memory, Knowledge, Wisdom, and Intelligence) enforce distinct persistence semantics: decay-based, supersession-based, evidence-gated, and session-scoped respectively. Write-time validation gates every entry by epistemic layer, preventing contradiction accumulation at the source rather than detecting it after retrieval. In a design validation on 500 annotated coding-agent sessions, CITE achieves 95% contradiction detection (vs 66% baseline) and 87% revision propagation (vs 12%).