Building a second brain that rewrites itself

For years my notes were write-only. I’d capture something, feel productive, and never look at it again. The problem wasn’t capture — it was that nothing ever changed after it landed. A note written in January stayed exactly as wrong in July.

So I rebuilt the whole thing around one rule: every new source must make the existing vault smarter, not just bigger.

The core idea

When a new source comes in, it doesn’t just get appended. It rewrites the pages it touches. If a fact changed, the old value isn’t deleted — it’s moved into a timeline with the date I learned it. That way I can ask who was my manager in February? and get a real answer.

The vault after an ingest should be different — not just longer. If old pages aren’t more connected and more current, the ingest wasn’t deep enough.

Notes written for a machine

The twist: I don’t really read these notes. Claude does. So every note is written AI-first — a plain-English preamble, rich frontmatter, a source link on every claim, and wikilinks everywhere so the graph stays traversable.

---
type: project
tags: [second-brain, systems]
ai-first: true
---

## For future Claude
A self-rewriting knowledge base. Read this
first to decide relevance in ~10 seconds.

Agents that work the night shift

Four scheduled jobs run on my work machine as plain launchd tasks: a morning daily note, a weekly health check, a weekly review, and a synthesis agent that hunts for patterns across unrelated sources. Each one runs Claude headless against the vault, writes its changes in place, and never commits — I review the diff over coffee.

• Nothing is deleted — only added, updated, or linked.
• Contradictions get flagged, not silently overwritten.
• Everything is logged, so I can audit what changed and why.

Measuring whether it actually works

A knowledge base you can’t retrieve from is a diary. So I built a small eval harness that generates questions from real notes and scores whether search finds the right one. The first baseline was humbling: the right note ranked #1 only about a third of the time, because long, chatty notes kept beating short canonical ones in a plain term-frequency ranking. Now retrieval quality is a number I can move, not a vibe.

It’s not magic. It’s just a filing cabinet that finally does its own filing — built on the open-source obsidian-second-brain framework, operated by Claude, and audited by me.