Hermes / NeuroArch ● wip
a self-evolving cognitive architecture
Hermes is the project I keep coming back to: a cognitive architecture that doesn't just answer prompts — it runs. It thinks in the background, keeps a persistent memory, and slowly reorganizes what it believes about the world.
Under the hood it's a market. Dozens of local models bid on cognitive work, and a router hands tasks to whoever's best for the job. Idle cycles aren't wasted — they go into research, belief recalibration, and a dream phase that crystallizes scattered notes into durable memory.
It's the closest thing I have to a system that's actually alive between conversations.
// highlights
- Idle cognition — keeps reasoning when nobody is watching
- Semantic memory with sentence-transformer retrieval (RAG)
- A "cognitive market" routing work across 30+ local models
- Dream-cycle crystallization of long-term memory
- DuckDB analytics engine + action-outcome learning
// stack
- Python
- DuckDB
- Ollama
- sentence-transformers
- RAG