From Agent Orchestration to Continuous-Time Cognition

Building a genuinely cognitive artificial system—one that perceives, acts, and accumulates structured experience through continuous engagement with its environment—is harder than it looks from the outside, and different in kind from building a system that scores well on benchmarks. This paper is an honest experience report: five generations of synthetic cognitive architecture, four of which were ultimately discarded. We trace the evolution from language model orchestration, through sequential processing pipelines and adaptive topological meshes, through Clifford algebraic fields, and finally to an exceptional Lie algebra substrate. For each generation we identify the motivating assumption, the failure mode that became apparent, and the precise lesson it forced. We synthesize eight transferable lessons. The central finding is counterintuitive: the right substrate for machine cognition is not a data structure, not a pipeline, and not a topological graph—it is a living algebra that accumulates operators rather than episodes, whose convergence under the Killing form constitutes a mathematically principled theory of memory consolidation.