An Argument for Externalized Epistemics

Where should an AI agent's knowledge live? Three options exist: in model weights (learned memory), in the context window (session-scoped), or in an external epistemic store (persistent, auditable). We argue that externalization is not merely convenient but architecturally required for agents that must maintain coherent beliefs over time, coordinate with other agents, or submit to audit. We present convergent evidence from five domains: information theory (neural networks have a capacity ceiling of ~3.6 bits per parameter), optimization theory (catastrophic forgetting is geometric, not incidental), interpretability (superposition and nonidentifiability make "what does the model believe?" ill-posed), distributed systems (multi-agent coordination requires shared substrate per CAP theorem), and neuroscience (biological memory externalizes episodic content before slow consolidation). We formalize these constraints as the Externalized Epistemics Thesis and present Layered Epistemic Agent Protocol (LeAP) as an architecture satisfying these requirements. CITE (Context In Tiered Epistemology) implements LeAP with four cognitive layers, write-time coherence enforcement, and AGM-compliant belief revision.