Built this after months of wrestling with AI context management.

Every chat system treats conversations as flat lists - works until topics branch mid-conversation. This routes messages to semantic branches (STAY=continue, ROUTE=reconnect, BRANCH=split), extracts facts with provenance, and assembles focused context for LLM calls.

20 messages instead of 100's. Sub-500ms.

Quick start: `make up` then `make dev` in the repo. Bring your own Groq API key.

Looking for feedback on the overall concept - is conversation branching solving a real problem or over-engineering? What would make this more useful?

Happy to answer any questions.