Darwin Gödel Machine

An open-ended, self-improving AI system that evolves its own source code using a local LLM. Built for autonomy, reflection, and code evolution — all running locally via Ollama and models like Llama 3.

Overview

The Darwin Gödel Machine (DGM) is a system that:

• Initializes with a basic coding agent
• Uses a local large language model (LLM) to propose self-modifications
• Evaluates modified agents via tests or benchmarks
• Retains improved agents in an evolving archive
• Runs completely offline with no API keys or cloud access

---

Features

• Fully local inference via Ollama + Llama3
• Code-editing agents powered by foundation models
• Archive of evolving agents (each with performance and full source)
• Open-ended exploration and branching from past agents
• Modular, inspectable components (tools, logs, test scoring)

---

Requirements

• Python 3.10+

• Ubuntu/Linux recommended

• Ollama installed and running locally

• llama3 or similar model pulled:

  ollama pull llama3

---

Setup

1. Clone the repo

   git clone https://github.com/yourusername/darwin-godel-machine.git
   cd darwin-godel-machine

2. Create virtual environment

   python3 -m venv venv
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Export environment variables (optional)

   export OLLAMA_MODEL=llama3
   export OLLAMA_API_URL=http://localhost:11434/v1

5. Run the Darwin Gödel Machine

   python dgm.py --benchmark_name SWE-bench --iterations 10

---

Project Structure

Darwin-Godel-Machine/
├── dgm.py                # Main DGM loop
├── coding_agent.py       # Agent logic (self-editing, benchmarking)
├── llm_withtools.py      # LLM interface (Ollama chat)
├── tools/                # Tool definitions (editor, bash, etc.)
├── utils/                # Logging, Git utils, scoring
├── archive/              # Stores all discovered agents
├── chat_history.md       # (Optional) conversation log
└── tests/                # Test infrastructure

---

Archive Output

Each run produces a growing archive/ of agents:

archive/
├── initial_agent/
│   ├── code/
│   └── performance.txt
├── child_1f43b2c3_perf_0.450/
└── child_a92fbbc2_perf_0.681/

Each subdirectory contains:

• The full working code of the agent
• Its performance score on the benchmark

---

Customization

• Modify evaluate_agent() in eval_utils.py to use your own tasks, metrics, or test suites.
• Add new tools in tools/ and expose them via prompt or callable interfaces.
• Replace the seed agent.py with your own scaffolding or language-specific agents.
• Extend scoring to include safety, diversity, or explainability.

---

Roadmap

• Long-term memory for agents
• Multi-agent competition
• UI for archive browsing
• Plugin-style tool framework
• Language support beyond Python

---

License

This project is licensed under the MIT License. See LICENSE for details.

---