Perdura
The mind that outlives its models.
A persistent knowledge graph where LLMs are ephemeral workers. Models board, contribute graph deltas, and leave, while the graph keeps a per-model track record of who to trust.
The idea
Multi-agent systems treat agents as permanent and memory as an afterthought. Perdura flips it: the memory is the persistent thing, and any model can be swapped mid-task with zero loss.
What I built
- Attributed graph where every node records which model wrote it
- Per-model reliability scorecards by domain
- A router that uses cheap local models by default and calls frontier models on a schedule
- Memoric binary: a 96-bit epistemic encoding for early contention detection
- Live dashboard and mind-map visualisation of each session
Built with
- Python
- SQLite
- Claude
- Gemini
- Qwen (LM Studio / Ollama)