What question did this study set out to answer?

The aim is to develop an operating system architecture that accommodates long-horizon reasoning for autonomous agents while maintaining context awareness.

April 3, 2026Open Access

SOMA: Sovereign Operating Memory Architecture

Key Points

The aim is to develop an operating system architecture that accommodates long-horizon reasoning for autonomous agents while maintaining context awareness.
Introduced a four-layer architecture consisting of working attention, episodic traces, sovereign knowledge, and environment grounding.
Developed a prompt-occupancy signal, Somatic Pressure (Pm), to regulate cognitive context management.
Created a reference runtime enabling self-regulated checkpointing and knowledge distillation.
Tested the architecture using extreme long-horizon tasks via the Leviathan benchmark series.
Demonstrated effectiveness in long-horizon tasks such as the TinyC Compiler and Chess Engine with successful performance under constrained Pm levels.
Showed that agents can manage their cognitive context autonomously without external intervention.

Abstract

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

Read Full Paperexternally

AI से पूछें

Bookmark

View Full Paper