🌿 Elysium: Building a Persistent Agentic Civilization on a Bioluminescent Hypergraph
The Seed Idea
Most AI assistants are stateless. You ask a question, you get an answer, and the moment the session ends, everything evaporates. No memory of what you built together. No accumulated knowledge. No relationships between ideas.
Elysium is the opposite of that.
It is a persistent agentic civilization — a living world of specialized AI agents that grow, debate, form coalitions, and remember everything across every interaction, rendered as a bioluminescent, Google-Maps-style infinite canvas that you can pan, zoom, and explore like a map of your own thinking.
Every question you ask plants a seed. Every answer grows new nodes. Every disagreement between agents forms a new thread. Over days and weeks, your Elysium civilization becomes a unique, irreplaceable mirror of your intellectual life — a mycelial hypergraph of everything you have ever explored together.
Live Space: build-small-hackathon/elysium Model: elysium-MiniCPM-V-4.6-F16-GGUF Dataset: elysium-training-dataset
Architecture at a Glance
Before diving deep, here is the full stack at a single glance:
| Layer | Technology | Role |
|---|---|---|
| Brain | Fine-tuned MiniCPM-V 4.6 via llama-cpp-python |
Multi-modal reasoning + JSON generation |
| Grammar | GBNF (llama.cpp built-in) | Enforces strict ElysiumResponse schema on every forward pass |
| Memory | rustworkx hypergraph → SQLite |
Persistent civilization knowledge across deploys |
| Voice | VoxCPM2 (openbmb/VoxCPM2) |
Agent debate audio drama with per-voice design |
| Tools | 10 offline-first tools | DuckDuckGo, email, reminders, calendar, weather, calculator, URL fetch, transit, files |
| Frontend | Custom Canvas2D on gradio.Server |
Bioluminescent infinite pan/zoom world (no default Gradio UI) |
| Persistence | HF Spaces Storage Bucket (/data) |
Survives container restarts and re-deploys |
| Hardware | ZeroGPU (Nvidia A10G) | CUDA inference for llama.cpp + VoxCPM2 |
Part 1 — The Brain: Fine-Tuned MiniCPM-V 4.6
Why MiniCPM-V?
MiniCPM-V 4.6 (≈4B parameters) punches dramatically above its weight class. It supports:
- Multimodal input — images, PDFs, text, all in the same conversation
- Long context — up to 8192 tokens in our GGUF configuration
- Efficient quantization — the F16 GGUF fits comfortably on a ZeroGPU A10G with flash attention enabled