Drag two conversations together. The model remembers everything.

A local AI chat interface where conversations are living objects — merge them, compile them into weights, and build a model that compounds knowledge over time. Fully local, fully yours, nothing leaves your machine.

What makes it different

Core concept

Most chat interfaces treat conversations as isolated sessions. CHATFORGE treats them as composable knowledge.

• Merge — drag any two conversations onto each other. Their contexts combine chronologically into a single conversation. Ask questions that require facts from both.
• Memory — after every exchange, the model silently extracts facts about you and stores them. Every future chat starts with that context already loaded.
• Compile (Phase 4) — one-click LoRA fine-tuning bakes merged context permanently into model weights. No context window needed at inference time.

Requirements

• Ollama with at least one model pulled
• Python 3.9+
• Node.js 18+
• NVIDIA GPU recommended (runs on CPU too, just slower)

Tested on RTX 3060 Ti with qwen2.5:7b / qwen3.5 — fits in 8GB VRAM.

Quick start

1. Clone

git clone https://github.com/gerritsxd/chatforge
cd chatforge

2. Backend

cd backend
pip install -r requirements.txt
python -m uvicorn main:app --reload --port 8000

3. Frontend

cd frontend
npm install
npm run dev

4. Open http://localhost:5173

Make sure Ollama is running with at least one model:

ollama pull qwen2.5:7b

Usage

Chat

Select a model from the dropdown, type a message. Conversations are auto-saved.

Merge

Drag one conversation from the sidebar onto another. They combine into a new MERGE: conversation. The model now has both contexts — ask questions that span both chats.

Memory

Click the MEMORY button in the topbar to see what the model has learned about you. Click × on any fact to make it forget. Memory is automatically injected into every new conversation.

Project structure

chatforge/
├── backend/
│   ├── main.py          # FastAPI server, chat streaming, memory extraction
│   ├── db.py            # SQLite: conversations, messages, memories
│   └── requirements.txt
├── frontend/
│   ├── src/
│   │   ├── App.jsx
│   │   └── components/
│   │       ├── Sidebar.jsx      # drag-to-merge lives here
│   │       ├── ConvItem.jsx     # draggable/droppable conversation item
│   │       ├── ChatWindow.jsx
│   │       └── MemoryPanel.jsx
│   └── package.json
├── start.bat            # Windows: launches both servers
└── PLAN.md              # full roadmap

Roadmap

• Phase 1 — Chat UI + Ollama streaming + SQLite persistence
• Phase 2 — Drag-to-merge conversations
• Phase 3 — Persistent memory with auto fact extraction
• Phase 4 — LoRA compilation: one-click bake context into weights
• Phase 5 — Docker container + hackathon demo polish

Stack

• Backend: Python, FastAPI, SQLite, httpx
• Frontend: React, Vite, dnd-kit
• AI: Ollama (any local model)

License

MIT