An Obsidian vault as a cross-tool AI external brain

Every AI tool wants to own your memory: chat histories here, "projects" there, none of it portable. My fix is one Obsidian vault — plain markdown, versioned with git — that acts as the external brain for all of them. No tool owns it; every tool reads it.

The honest part most setups skip: there are two kinds of tools, two different realities.

• Local coding agents (Claude Code, Codex) have filesystem access. They read the vault's CLAUDE.md contract on boot, load the relevant project note and the latest daily log, and write their own session notes as they work. Fully automatic.
• Web chats cannot see local files, and no magic sync pretends otherwise. The bridge is a cheap ritual: paste a session-start block going in, paste a handoff prompt going out — the model emits a daily-note entry in vault format, and I commit it.

The structure is deliberately thin: one master note per project, reusable runbooks for repeated jobs, dated daily logs, and an append-only decisions.md with the why behind every choice. A handful of rules keep it alive:

1. Read before acting — project master first, then the latest daily.
2. Append, don't proliferate — update an existing note unless the topic is genuinely new. Bloat is the failure mode that kills these systems.
3. A daily note per real session — what happened, what was decided, what's next.
4. Log decisions with reasons — future-me and future-agents search these constantly.

The part I underestimated: useful notes are the residue of real work, not something you design upfront. Two apps, one Neon database exists because an agent and I nearly learned it the hard way, and the vault was where the lesson landed. This garden is just the vault's public face — notes that earned a publish: true. See Why this garden exists for that half of the story.