Version: 1.0
Date: April 2026
Author: Ilja Schots
Co-design & implementation: Mira (Claude, Anthropic)
License: Custom (free for personal/research use; commercial use requires license)

Gemma is an experimental cognitive architecture that wraps a local large language model (running via Ollama) in five processing layers, transforming a stateless text generator into a system with persistent memory, self-monitoring, background reflection, and a continuously updated self-model. The base LLM remains unmodified — it serves as "raw instinct" (Layer 1), while the surrounding layers add continuity, error correction, and self-awareness.

The entire system runs offline on consumer hardware. No API keys, no cloud services, no external dependencies beyond the model weights.

Layer 1:   Instinct         — Local LLM via Ollama (raw generation)
Layer 1.5: Syntactic Anchor — Surface guard (pre-Monitor, zero LLM calls)
Layer 2:   Monitor          — Semantic evaluation before output delivery
Layer 3:   Episodic Memory  — Weighted graph of experiences (persistent, on disk)
Layer 4:   Background       — Between-session reflection (separate process)
Layer 5:   Self-Model       — Persistent self-representation (JSON, updated by the system)

Input
  → Layer 1 generates raw response
    → Layer 1.5 checks syntactic integrity (zero-cost, no LLM call)
      → Layer 2 evaluates semantics (consulting Layer 5 self-model)
        → Output delivered to user
          → Layer 3 records the exchange as episodic memory
            → Layer 4 reflects between sessions (background process)
              → Layers 3 + 5 updated with new insights
                → Next session informed by accumulated history

The LLM is instinct, not intelligence. A single LLM call is fast but has no continuity, no self-knowledge, and no error-correction capability. The five layers add these without modifying the model itself.

The Self-Model is functional, not decorative. The Monitor (Layer 2) actively consults the Self-Model before evaluating output. The Background Processor rewrites it between sessions based on observed patterns.

The Syntactic Anchor catches what the Monitor misses. Added after observing syntactic drift in the base model under high load. Purely structural checks — well-formed tokens, balanced brackets, degenerate output detection — at zero computational cost.

Reflexive Tuning closes the loop. The Monitor observes output quality over time and feeds temperature adjustments back to Layer 1. The system tunes its own generation parameters.

Hardware awareness enables graceful degradation. GPU temperature, VRAM usage, and CPU load are monitored. Under thermal pressure, expensive operations are suspended to prevent drift.

• Python 3.10+
• Ollama running locally
• A model pulled in Ollama (default: gemma4:26b, configurable)
• Python packages: requests, pyyaml, psutil

pip install requests pyyaml psutil

Recommended hardware: 24+ GB VRAM for the 26B model. Smaller models are supported and configurable.

# Interactive CLI
python3 gemma_engine.py

# HTTP server (OpenAI-compatible API)
python3 gemma_server.py

# Background reflection process
python3 background_processor.py

Edit gemma_identity.yaml to set identity, personality, and generation parameters (temperature, top_k, top_p). The included file is an example — replace the placeholder text with your own configuration.

The self_model.json is created and maintained by the system at runtime. You can read it, but don't hand-edit it.

• Not a fine-tuned model. The base LLM is unmodified.
• Not a multi-agent framework. One model, multiple processing layers.
• Not production software. This is research infrastructure.
• Not safe to expose to the internet without authentication.

These files are generated automatically when the system runs:

• self_model.json — the system's self-representation
• episodic_memory.db — SQLite database of experiences
• self_model_checkpoints/ — versioned self-model snapshots
• conversations/ — conversation logs
• reflections/ — Layer 4 reflection outputs

@software{schots_gemma_2026,
  author       = {Schots, Ilja},
  title        = {Gemma: A Five-Layer Cognitive Architecture for Local LLMs},
  year         = {2026},
  month        = {4},
  publisher    = {Zenodo},
  doi          = {[DOI assigned at upload]}
}

Free for personal, educational, and research use. Commercial use requires a separate license — see LICENSE file for details.