second-brain-template

An Obsidian vault that your coding agent runs for you — Claude Code, Codex, or any agent that reads AGENTS.md — built so that answering from your notes costs a few hundred tokens instead of all of them.

What it is

You capture anything — quick notes, a daily log, documents, ideas, drawings — and your agent compiles it into a knowledge graph of small, linked notes. Every note leads with a capped answer surface and keeps its full detail below it, and a generated index lets a question reach the right lines without reading anything else. Capture can be slow and thorough; answering stays cheap.

The claim it defends: answering costs 30–50× less than loading your notes, and nothing is summarised away to get there. In the vault this method was built in, the route-and-card lookup measures a median of about 44× cheaper than loading its ~16,800 tokens of sources. On the shared fictional benchmark in bench/, the first scorecard is 27× on an ~11,000-token corpus, where a smaller corpus lowers the ratio for the same answer.

The drawing as text, and what each stage does: docs/architecture.md.

Three ways in

Convert the vault you already have — the usual case. Download or clone this repository anywhere, open your coding agent in your vault's folder, and say:

Read /SETUP.md and set up my vault.

It backs your vault up first, deletes none of your notes, and works only in your vault — it never changes this folder.