From Memory to Epistemics

Where should an AI agent’s knowledge live?

In the weights? Train on your data, encode it in parameters. But weights have a capacity ceiling (~3.6 bits/param), forgetting is geometric, and you cannot audit what the model believes.

In the context window? Pass everything at inference time. But context is session-scoped. Nothing persists.

In an external store? This is RAG, but RAG barely scratches the surface. It externalizes content without externalizing epistemic structure. Every stored fact has the same standing, contradictions accumulate, and corrections do not propagate.

We argue externalization is not optional. It is architecturally required.

Evidence converges from five domains: information theory (capacity limits), optimization (catastrophic forgetting), interpretability (beliefs are non-auditable in weights), distributed systems (multi-agent coherence needs shared substrate), and neuroscience (even brains externalize before consolidating).

In-weights fails on capacity, durability, and auditability. Per-agent learned state fails on coordination. Session context fails on persistence. External epistemic state is what remains.

The question is not whether to externalize. The question is what structure the external store should have. The paper makes the case and sketches what that structure looks like.

Coming soon, follow progress at github.com/engrammic-ai/research →