Canonics: Versioning the Frame, Not the Facts — Vigency in AI Memory After Judgment Expires

A stored interpretation can stay factually true and still stop being valid guidance. Current knowledge-management systems demand exhausting vigilance: each new fact may require manual changes across many derived notes, plans, and summaries. Long-lived AI agents then fail by retrieving clean, sourced memory from a regime that no longer governs the decision at hand.

Canonics names and manages that missing state: canonical vigency, whether an interpretation is still in force under the active canon. The framework adds two primitives. A canon event is a human-declared regime change with explicit scope; an epoch is the per-domain regime interval derived from those events. Understandings carry their epoch, source provenance, and canon distance (κ), the count of in-scope canon events crossed since derivation. Facts stay immutable. Interpretations become versioned cache. Retrieval can then separate current guidance, historical belief, material needing re-derivation, and revoked guidance.

The paper gives the ontology, operations, integrity model, Canon Review protocol, and evaluation agenda. The practical entry point is modest: a current-regime block plus epoch_id and canon_distance tags. Canonics does not detect shifts by itself. It lets operators declare them once, propagate the blast radius, and audit what old memory is still allowed to do.