From Stateless Models to Stateful Intelligence - A Systems Hypothesis for Long-Horizon Artificial Intelligence

Abstract Over the past several years, advances in large language models have significantly improvedreasoning quality, planning ability, tool use, and agentic behavior. Contemporary AI systemsare increasingly capable of solving complex tasks within a single reasoning episode. Yetdespite these advances, most systems continue to reconstruct their cognitive state wheneverreasoning resumes after an interruption. Context may be retrieved, memories may besearched, and previous conversations may be summarized, but the underlying cognitiveprocess effectively begins anew. This paper argues that this recurring reconstruction may represent a broader architecturallimitation rather than merely a consequence of finite context windows. As AI systemsbecome expected to participate in projects that unfold over days, months, or years, the abilityto preserve and evolve cognitive state may become as important as the ability to reasonwithin a single interaction. The central hypothesis of this paper is that persistent cognition should be treated as a firstclass systems abstraction, distinct from memory storage, retrieval mechanisms, knowledgerepresentation, or model parameters. Rather than viewing intelligence as isolated reasoningepisodes connected only through retrieved information, this paper proposes an architecturallayer responsible for preserving, organizing, projecting, and evolving cognitive state acrossdiscontinuous reasoning sessions. To explore this hypothesis, the paper introduces a conceptual decomposition consisting of theCognitive Commit Protocol (CCP), Cognitive Graph Protocol (CGP), CognitiveTraversal Protocol (CTP), and an orchestrating Cognitive Operating System (COS).Together, these components are presented not as a production architecture, but as onepossible realization of a broader systems abstraction. This document is intentionally framed as a position paper. Its purpose is to establish aresearch direction, articulate a systems hypothesis, and invite technical discussion regardingwhether persistent cognition represents a missing architectural layer for long-horizonartificial intelligence.