🧠 Jonas

Local-first AI assistant with a neural policy brain.

Jonas is not just a chatbot wrapper around an LLM.

He separates:

🗣 Language Model → speaks

🧠 Neural Policy Brain → decides how he speaks

🗂 Semantic Memory (SQLite + embeddings) → remembers context

📊 Online Feedback Loop → learns from your ratings

Fully local. No cloud APIs. Runs clean on Apple Silicon.

🔥 Philosophy

Most AI assistants are just prompt wrappers.

Jonas is different.

Instead of letting the LLM decide everything, Jonas uses a small PyTorch neural network to:

Adjust temperature dynamically

Control verbosity

Inject subtle humor

Decide how much memory to recall

Learn from /rate feedback

The LLM generates language.

The policy brain shapes behavior.

🏗 Architecture User Input ↓ Embedding (local via Ollama) ↓ Neural Policy Brain (PyTorch MLP) ↓ Memory Recall (SQLite + cosine similarity) ↓ LLM (Ollama) ↓ Response ↓ Optional /rate feedback → trains policy 🎭 Personality

Jonas carries:

Tactical calm operator energy

Intelligent older-brother warmth

Quiet strategic executive clarity

Subtle, controlled wit

Grounded. Culturally fluent. Never theatrical. Never dismissive.

⚙️ Setup (Mac / Apple Silicon Recommended) 1️⃣ Install Ollama brew install ollama ollama serve 2️⃣ Pull models

Recommended:

ollama pull llama3.1:8b ollama pull nomic-embed-text

(You can use llama3.2:1b for lightweight testing.)

3️⃣ Python Setup python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt 4️⃣ Run Jonas python -m src.main 💬 Commands Command Description /quit Exit /reset Clear memory /history Show recent turns /rate 0.8 Train Jonas on last response 🧠 Neural Policy Brain

The policy network:

Small PyTorch MLP

Learns from scalar reward (0–1)

Saves weights to data/policy.pt

Adapts tone and structure over time

No reinforcement learning frameworks. No heavy infrastructure. Just incremental behavioral shaping.

🔒 Privacy

All inference is local via Ollama.

Memory stored in local SQLite database.

No external API calls.

No telemetry.

📁 Project Structure jonas/ │ ├── src/ │ ├── main.py │ ├── policy_brain.py │ ├── persona.py │ ├── encoder.py │ ├── memory.py │ ├── ollama_client.py │ ├── data/ │ ├── chat.db │ ├── policy.pt │ ├── requirements.txt └── README.md 🚀 Roadmap

Planned improvements:

Policy smoothing (reduce tone switching)

Working-memory summarization

Structured task mode

Optional lightweight GUI

Long-term preference tagging

Better reward shaping

🧩 Why Jonas Exists

To experiment with:

Policy steering for LLMs

Local-first AI systems

Minimal RL-style feedback loops

Human-in-the-loop tone adaptation

Jonas is a research playground disguised as a personal assistant.

⚠️ Notes

Smaller models (1B) may exaggerate tone.

8B models handle nuance significantly better.

Delete data/policy.pt if you change embedding size.

🪶 License

MIT