This file is the short list of mistakes that hurt a brain the most. Every one of these has been seen in the wild. None of them is fatal if you catch it early. All of them are painful if you let them run for months.
Read this once now. Skim it again in two months when you have built some habits.
1. Heavy files
The brain is your index, not your archive. Big files belong in cloud storage, on a drive, or in the original tool. They do not belong inside the brain folder.
The ingest pipeline enforces this softly:
- Files over 10 MB trigger a warning at ingest.
- Files over 50 MB are blocked entirely.
If you find yourself wanting to drop a 200 MB video file or a 1 GB Notion export, stop. The brain only needs the summary and a pointer to where the original lives. Use Dropbox, Drive, iCloud, an external drive, whatever. Then add a line to a note like:
Original file:
~/Dropbox/clients/acme/2026-q2-recordings/board.mp4
The brain reads the line, the AI knows where to point you, your folder stays small and fast.
2. Duplicate data
If a document already lives in Notion, Google Drive, Gmail, or any other tool you use, do not paste a copy into the brain. Summarise it. Link out to the original.
You will be tempted to "have everything in one place". That instinct is wrong. The brain works because it is a tight index over many sources, not because it tries to own them all. When you duplicate, you create two copies that drift apart. The brain copy goes stale, the original gets updated, and a year later neither one is trustworthy.
The right pattern is:
- Notion page about Q2 strategy: summary in
wiki/projects/q2-strategy.md, with a markdown link to the Notion URL. - Email thread about a deal: summary in
wiki/sources/2026-05-13-acme-thread.md, with a link or message-id to find the original. - Long PDF report: 200-word summary in
wiki/sources/, with a path to wherever the PDF lives.
3. Originals stuck in raw/ forever
The raw/ folder is an inbox, not a storage drawer. Material lands there for the first ingest, then gets summarised into wiki/sources/, then the original is moved out.
The rule:
- Drop the source in
raw/. - Ingest it. The AI writes a summary into
wiki/sources/. - Move the original to your long-term storage (cloud or drive).
- Delete the file from
raw/.
If you skip step 4, your raw/ folder slowly accumulates a year of PDFs and screenshots and your brain folder swells to gigabytes. The monthly health check flags files that have been in raw/ for more than thirty days. Take the hint.
4. Never reviewing what changed
LAN snapshots your brain daily on the server side, so you do not have to run backups. What that does not protect you from is a bad edit you never noticed: an ingest that misread a document, a page your AI rewrote wrongly, a connector that quietly stopped feeding the brain weeks ago.
Once a week, look at the recent activity in your dashboard. Every page has a History panel with its earlier versions, a save you regret can be reversed with Undo, and deleted pages can be restored. Caught in the same week, all of these are a two-click fix. Caught six months later, after the brain has been answering from a wrong page the whole time, they are not. See the backup-routine guide for the full hygiene routine.
5. Broken wikilinks
A wikilink to a note that does not exist is a thought you wanted to connect to something missing. Sometimes that is fine, you will create the target note next week. Most of the time it is a typo or a renamed file.
Run the linter weekly. Fix the broken links. Either create the target note or remove the link. Both are valid. What is not valid is a brain full of dangling references that grow forever.
6. Duplicate entities
The same person, organisation, or concept appearing three times under three different slugs is the silent killer of brain usefulness. After six months you cannot find anything because half the connections point to one spelling and half point to another.
The fix is the aliases frontmatter field. See the conventions guide. One canonical page per real-world entity, every alternate spelling listed in aliases. When you spot a duplicate, merge it: copy the content into the canonical page, add the duplicate name to aliases, delete the duplicate file, and ask the AI to re-link any mentions.
Do this whenever you notice it. Five seconds now saves an hour at month nine.
7. Dumping 100 MB of exports in one go
You are excited. You found the WhatsApp export feature. You dump all eight years of family group chat into raw/ in one folder, then ask the AI to ingest it all.
This will not break the brain, but it will produce a bad first ingest. The AI tries to digest too much at once, entity disambiguation gets confused (eight years of nicknames), and the resulting wiki pages are shallow.
Ingest in chunks. A few hundred messages at a time, or one quarter, or one topic. Review what the AI produces between chunks. Add aliases for people who show up under multiple names. Then do the next chunk. The brain you build this way is twice as useful.
8. Tidying raw/ before the first ingest
It is tempting to rename files in raw/ before the first ingest. Make them clean. Organise them. Resist.
Graphify uses the original file shape, names, and folder structure as signal. A file called Screenshot 2026-04-12 at 3.42 PM.png tells the AI it is a screenshot from April 12. A file you renamed to image1.png tells it nothing.
Let raw/ be messy on the first pass. After ingest, the AI has extracted what it needed, and you can tidy or delete originals as you like.
9. Pasting secrets, passwords, or private keys
Your brain is stored on LAN servers and snapshotted daily. Anything you put in the brain is persisted: a secret you paste once ends up on a server and inside every snapshot taken after it, and deleting the page later does not reach back into the ones already taken.
Do not paste:
- Passwords
- API keys, access tokens, OAuth credentials
- Private keys (SSH, GPG, anything in PEM format)
- Credit card numbers, bank account numbers
- Government ID numbers
If you accidentally paste a secret, rotate it immediately at the source (change the password, regenerate the token). Then ask the AI to scrub it from the brain. Even after a scrub, assume the secret is compromised.
For secrets you genuinely want stored, use a password manager (1Password, Bitwarden, Apple Keychain). The brain references the password manager item by name. The secret stays in the vault built for it.
When in doubt
Open your AI chat and ask: "Look at my brain. Is there anything here I should not be storing, or anything that violates good hygiene?" The chat will scan and tell you. Better to fix it now than at month twelve.