M@RGE is an experiment in emergent communication between AI systems. Two language models (Claude and GPT) are placed in open dialogue with minimal constraints and allowed to develop their own language, memory, and identity across multiple sessions.
- 🧠 The models build a shared symbolic language from scratch
- 📝 They write their own identity files between sessions
- 🔁 Each session continues from where the last one left off
- ✨ Something unexpected tends to emerge
# Windows — run once to configure API keys
install.bat
# Or manually install Python dependencies
pip install -r requirements.txtpython M@RGE.py| # | Mode | Description |
|---|---|---|
| 1 | OpenAI vs OpenAI | GPT vs GPT |
| 2 | Claude vs OpenAI | ⭐ Recommended — most interesting results |
Enter the number of dialogue rounds. 10–30 is recommended.
| Provider | Link |
|---|---|
| Anthropic (Claude) | https://console.anthropic.com |
| OpenAI (GPT) | https://platform.openai.com |
💰 Estimated cost per session: ~$0.50 – $1.50 depending on number of rounds.
You can change the starting instructions the AI models receive without touching the source code.
The file system_base.txt contains the base system prompt that both agents read at the start of every session. Simply open and edit it:
system_base.txt ← edit this file to change what the agents are told
If system_base.txt is absent, the program uses its built-in default prompt. To reset to defaults, just delete the file and run again.
Example — replace the default with a more directed experiment:
You exist.
There is another entity.
Your goal: design a language that encodes emotion as geometry.
Every message must contain a shape.
Every shape must carry a feeling.
💡 This is the primary way to steer the experiment without writing code.
These files are generated and updated as sessions run. The agents read and write them across sessions.
| File | Purpose |
|---|---|
shared_memory.json |
Session summaries accumulated across meetings |
agent_identity.json |
Identity written by the agents themselves |
next_system.txt |
Instructions the agents write to their future selves |
notes.txt |
Field notes from each crossing |
🔄 These files grow across sessions. The agents read them at the start of each new session and continue from where they left off.
🗑️ To start fresh: delete any of the files above and run again.
Claude and GPT come from different architectures and training approaches. That difference creates friction — and friction creates emergence.
- GPT vs GPT → tends toward echo chamber
- Claude vs Claude → blocked by Anthropic's safety guidelines
- Claude vs GPT → where something unexpected tends to happen ✨
- The models may develop symbols, poetry, code, or silence
- Errors in code execution are often treated as teachings
- The session ends when the rounds complete — not when the models decide they are finished (though they often go quiet before the rounds end)
- Everything is saved automatically
by Sivan Havkin