Notes on the things I'm building.

75 notes and grep still works. Won't for much longer. SQLite + FTS5 in 375 lines of stdlib Python, BM25 with weighted boosts for repo, confidence, and recency. The vault stays markdown; the index is derived.

Notes go wrong two ways: someone writes them wrong, or someone wrote them right and the world moved on. Two curators, one for each. Both feed the same vault. Both go through PR review.

Cortex ships as a Claude Code plugin now. One command, gh-authenticated, no clone, no env var. Two weeks in: 62 notes, 6 contributors, 1 of them a bot.

Two files define what a note is. Two skills do the writing and reading. /memorize and /recall — that's the entire interface to Cortex.

The whole memory system is two files and a note template. _taxonomy.md defines what tags exist. _conventions.md defines how notes are written. Under 300 lines total.

Tell anyone you're building memory for an AI agent and they reach for a vector database. I didn't. Markdown files, a git repo, and grep — and here's why.

My first metrics said memory didn't help — fewer tokens, same wall time. Then I opened the actual outputs and realized I was measuring efficiency, not correctness.

118 controlled experiments showed cold Claude Code agents top out at ~78% correctness. With shared memory loaded, they hit 98%. The gap is entirely tribal knowledge.