SOMA: Sovereign Operative Memory Architecture

A cognitive operating system for long-horizon autonomous agents

---

📖 Overview

SOMA (Sovereign Operative Memory Architecture) is a four-layer cognitive architecture for autonomous agents that enables long-horizon task completion through:

• L1: Bounded working attention with explicit context management
• L2: Immutable episodic record (audit trail)
• L3: Sovereign persistent knowledge (cross-session memory)
• L4: Environment grounding (IDE, browser, terminal, git)

Key innovation: Somatic Pressure ($P_m$), an explicit prompt-occupancy signal that allows the agent to self-regulate context saturation.

🏆 Benchmarks

Leviathan benchmarks demonstrate SOMA's capabilities on extreme long-horizon tasks:

Task	Turns	Peak $P_m$	Result
TinyC Compiler (from scratch)	132	<20%	✅ Pass
Chess Engine (Perft verified)	99	<20%	✅ Pass

Both completed on free-tier LLM (Gemini Flash Lite) with zero monetary cost.

🚀 Quick Start

Try soma-lite (Reference Implementation)

The minimal reference implementation (~700 LOC, zero dependencies):

npm install soma-lite

Or clone directly:

git clone https://github.com/mcarbonell/soma-lite.git
cd soma-lite
node run-agent-lite.js

👉 Documentation: soma-lite README

📚 Documentation

Academic Paper

• Paper: docs/research/SOMA-Paper.pdf
• Source: docs/research/SOMA-Paper.tex
• Citation: See CITATION.cff

Research Notes

Deep dives into specific research directions:

• Mutable Working Memory: docs/research/research-mutable-scratchpad-l1.md
• Synthetic Hippocampus: docs/research/SOMA_SYNTHETIC_HIPPOCAMPUS_DESIGN.md
• Auto-LoRA Consolidation: docs/research/research-auto-lora-consolidation.md
• Sovereign Kernel v9: docs/research/research-soma-v9-sovereign-kernel.md
• Tensor-Level Injection: docs/research/research-tensor-level-memory-injection.md

Benchmarks

• Leviathan Suite: benchmarks/run-benchmarks.js
• TinyC Compiler: Full specification and evaluator
• Chess Engine: Perft verification (depth 1-5)

🧠 Architecture

┌─────────────────────────────────────────┐
│  L1: Bounded Working Attention          │
│  ├─ Structured prompt (task + panels)   │
│  ├─ Pm dashboard (context pressure)     │
│  └─ Agent-controlled reconstruction     │
└─────────────────────────────────────────┘
              ↕ checkpoint/distill
┌─────────────────────────────────────────┐
│  L2: Immutable Episodic Record          │
│  ├─ JSONL action ledger                 │
│  ├─ Milestone history                   │
│  └─ Session-scoped audit trail          │
└─────────────────────────────────────────┘
              ↕ retrieve/promote
┌─────────────────────────────────────────┐
│  L3: Sovereign Knowledge                │
│  ├─ Identity template                   │
│  ├─ Semantic memory (embeddings)        │
│  ├─ File annotations                    │
│  └─ Knowledge tree                      │
└─────────────────────────────────────────┘
              ↕ observe/act
┌─────────────────────────────────────────┐
│  L4: Environment (IDE Suite)            │
│  ├─ File tree + editor                  │
│  ├─ Terminal panel                      │
│  ├─ Browser panel                       │
│  └─ Git panel                           │
└─────────────────────────────────────────┘

🔬 Research Directions

SOMA extends beyond the current implementation with six forward-looking directions:

1. Mutable Working Memory — Editable scratchpad during inference
2. Synthetic Hippocampus — System 1 network for sub-millisecond intuition
3. Autonomous LoRA Consolidation — Overnight fine-tuning from experience
4. Sovereign Inference Kernel — Tensor-level memory injection
5. The Genesis Chain — Self-bootstrapped engineering feat (8 stages)
6. Self-Authored Identity — Agent chooses its own frame of reference

See docs/research/ for detailed specifications.

🎯 Philosophy

"More context is not equivalent to better cognition."

SOMA is inspired by Sutton's Bitter Lesson: leveraging computation and learning at scale consistently outperforms methods that encode human knowledge.

Applied to agent orchestration:

• ❌ Hard-coded heuristics in orchestrator = ceiling
• ✅ Model's own reasoning with right substrates = scales with compute

🤝 Contributing

This is primarily a research project. Contributions welcome in:

• Bug reports and fixes
• Benchmark improvements
• Research discussions
• Extension implementations

Please open an issue before starting a PR to discuss scope and design.

📄 License

• SOMA documentation & paper: CC BY 4.0
• soma-lite implementation: MIT License
• Research notes: Open for discussion

👨‍💻 Author

Mario Raúl Carbonell Martínez

• GitHub: @mcarbonell
• Email: mrcm.10@outlook.es

📬 Citation

If you use SOMA in your research, please cite:

@article{carbonell2026soma,
  title={SOMA: Sovereign Operative Memory Architecture},
  author={Carbonell Mart{\'i}nez, Mario Ra{\'u}l},
  journal={arXiv preprint arXiv:2603.XXXXX},
  year={2026}
}

See CITATION.cff for full citation metadata.