C2Cognitive Core: Evidence-Bounded Persistent Cognition for AI Coding Agents

Abstract

Persistent cognition changes the assurance requirements of AI coding agents because durable memories, lessons, reusable skills, structural observations, and derived knowledge can influence later engineering work after repository state, task context, or execution sessions have changed. C2Cognitive Core v1.0.0 presents a repository-resident cognitive-engineering system designed to preserve useful cognition without allowing persistence itself to become repository authority. The system separates physical ingestion, evidence admission, persistent cognition, continuity state, model-facing representation, and repository effects. Bounded Read v2 (BR-v2) constrains physical ingestion before semantic selection. Governed admission evaluates path containment, sensitivity policy, provenance, integrity, freshness, visibility, and applicable authority before information can influence durable cognitive state. Admitted cognition is represented through typed L0-L3 Memory, advisory Skills, source-bound Structural Candidates, derived Wiki state, and bounded Agent Loadouts. Handoff, checkpoint, and resume artifacts preserve work continuity but remain distinct from cognitive truth. Adaptive Context Representation Planning (ACRP) may change representation only after the semantic evidence set has been selected and frozen. Model, cache, routing, and adapter telemetry do not become evidence or write authority. Model and worker outputs remain proposals until current Goal state, lease/fence conditions, exact write scope, rollback basis, and other applicable authorization are revalidated near effect time. C2ModelAdapter v0.5.5 is kept at the host/runtime boundary rather than promoted into persistent cognition. The v1.0.0 release verification includes 59/59 registered checks for each English and Indonesian edition, 71/71 Core regression tests, the 74-test C2ModelAdapter collection, 30,005/30,005 router-combination simulations, 50,001/50,001 effective-route simulations, and a 524,288-state finite progress-liveness analysis with no invariant violation in the declared model. These results are bounded evidence for the executed surfaces and are not a claim of universal correctness, security, provider behavior, or defect absence. C2Cognitive's contribution is a practical cognitive-engineering discipline for making durable agent cognition explicit, inspectable, provenance-bound, freshness-aware, resumable, and authority-bounded during long-horizon AI-assisted software engineering.