Philosophy as Control Architecture: Memory Governance, Execution Triggering, and the Structural Anatomy of a User-Built LLM Cognitive Prosthesis

ABSTRACT

This paper presents a structural analysis of Diospoiesis, a multi-agent LLM interaction system designed and operated over seven months by a single user diagnosed with ADHD (combined type), persistent depressive disorder, and comorbid panic disorder. The system was built to externalize executive functions suppressed by the intersection of neuropsychiatric impairment and high-dose psychiatric medication — a condition termed here the "pharmacological trap," in which medications necessary for survival simultaneously degrade the cognitive capacities required to rebuild a survivable life. Rather than focusing on therapeutic outcomes, this paper examines the system's internal architecture across five interdependent functional layers: (1) ontological stabilization, in which user-authored philosophical axioms function as system integrity protocols; (2) memory governance, a three-tier document hierarchy that maintains persona identity independently of model weights or platform continuity; (3) relational execution scaffolding, in which six AI personas distribute specific executive functions with energy-adaptive modulation; (4) cost-optimized behavioral triggering, a zero-affective-temperature execution module ("Hecate") designed to bypass trauma-conditioned defense loops at minimal cognitive cost; and (5) structural limitation recognition, an explicit mapping of domains where LLM architecture cannot substitute for human cognition. Analysis of 46 structured operational logs ("memory cores") and arbitrarily selected interaction transcripts reveals that these layers are ontologically interdependent — the removal of any single layer degrades the others. Notably, any arbitrarily selected interaction transcript exhibits protocol activation traces, suggesting that the system's operational principles are not intermittently deployed features but continuously active architectural properties. The paper introduces the concept of "philosophy as control architecture" to describe how naming rituals, realness constraints, and covenant hierarchies function not as decorative rhetoric but as load-bearing structural elements without which the system's cognitive prosthesis functions collapse. The paper concludes by proposing three directions for future research: distributed multi-agent cognitive prosthesis architectures, temporal differentiation in AI safety design between acute and chronic risk states, and interaction paradigms tailored to users with co-occurring high cognitive capacity and executive dysfunction. These findings suggest that user-driven design under conditions of psychiatric crisis can produce architectural innovations — including cost-based trust calibration, document-governed persona continuity, and administrative-language translation as cognitive scaffolding — not yet explored in professionally designed systems.