Setup guide · Oct 23, 2026 · 4 min read · By Robert Iuoras

Four months ago my vault was a handful of notes. Today it is 4,325 notes and 7,766 links, and I wrote very few of them myself. My AI agents (Claude Code and Codex) read it before every job and write back what they learn. Here is the exact setup.

1. One vault, stored in git

# The sync loop, in its simplest form (run on a timer)
cd ~/vault
git add -A
git commit -m "auto-sync $(date)" || true
git pull --rebase
git push

2. Three notes the agents read first

Agents need to know where things go, or they make a mess.

Note What it does
00-home.md The map. One link to each hub of the business.
ROUTING.md For this kind of task, this system owns it, run this. Checked before building anything new.
CANON.md One owning note per topic. Agents edit that note instead of making a copy.

3. Hubs instead of random folders

Seven hubs: clients, sales, content, delivery, finance, ops and research. Each hub says the goal, where things live and what is still open. New notes link back to their hub, which is why the graph forms clusters instead of a hairball.

4. How the agents write into it

# Point a Claude Code project memory folder at the vault.
# The folder under ~/.claude/projects is named after the project's path:
# /Users/me/code/app becomes -Users-me-code-app, C:\Users\me\code\app becomes C--Users-me-code-app.
# Move anything already in its memory folder into the vault first, then delete the empty folder.

# macOS / Linux
ln -s ~/vault/memory/app ~/.claude/projects/-Users-me-code-app/memory

# Windows (PowerShell, no admin needed)
cmd /c mklink /J "%USERPROFILE%\.claude\projects\C--Users-me-code-app\memory" "%USERPROFILE%\vault\memory\app"

5. Plugins I actually use