Marrow
Everything you know, one mention away.
StatusBuilt · 2026
Sole author
The problem
Notes, decisions and links end up in ten places, and search only finds the exact words.
What I built
- A personal second brain: one SQLite core with two doors, MCP for Claude and REST for the web UI.
- Hybrid recall: keyword search and a local vector model fused with Reciprocal Rank Fusion, so “car insurance” finds “auto coverage renewal”.
- Answers with quote-anchored citations, and a non-AI fallback for every AI step.
How it stays safe
- Every answer cites the notes it came from.
- Every AI step has a non-AI fallback.
- This page shows a seeded demo brain, not real notes.
Anatomy
- One SQLite core
- Topics that keep their place
- Ask in plain words
- Found by meaning
- Quote-anchored citation
Numbers
- 49commits in four days
- 7MCP tools for Claude
- 95unit tests
Stack
- TypeScript
- SQLite
- MCP
- Claude
Sources
- From the project’s repository
- About this website