SOMA: Sovereign Operating Memory Architecture

Abstract Large language models are effective local reasoners but poor long-horizon operators. In demanding environments such as software engineering, compilation, or game-playing, the dominant failure mode is not insufficient intelligence but the absence of bounded, persistent, and inspectable cognitive state. This paper introduces SOMA (Sovereign Operating Memory Architecture), a four-layer cognitive operating system for autonomous agents. SOMA separates bounded working attention (L1), immutable episodic traces (L2), sovereign persistent knowledge (L3), and environment grounding (L4). The architecture makes context saturation explicit through a prompt-occupancy signal called Somatic Pressure (Pm), and gives the agent direct metacognitive affordances to checkpoint, distill, and reorganize its own working state without external intervention. Key Contributions - Four-layer memory and environment architecture for long-horizon agents- Somatic Pressure (Pm), an explicit prompt-occupancy signal for self-regulated context management- Reference runtime with self-regulated checkpointing and distillation- Leviathan benchmark series: extreme long-horizon tasks with deterministic evaluators- Six research directions for future work Benchmarks - TinyC Compiler: 132 turns, Pm < 20%- Chess Engine (Perft verified): 99 turns, Pm < 20%- Both on free-tier LLM with zero monetary cost Links - GitHub: https://github.com/mcarbonell/soma- soma-lite (npm): https://www.npmjs.com/package/soma-lite- soma-lite (GitHub): https://github.com/mcarbonell/soma-lite