OzBrain vs a DIY LLM wiki
Updated
The short answer
OzBrain is the hosted version of the LLM wiki pattern: a structured knowledge base your agents read, write, and keep coherent, behind one connector URL, shareable with the people you work with. If you built the do-it-yourself version after reading the gist, you already believe in the pattern. What separates the two is who does the librarian work, and how far past your laptop the wiki reaches.
You also already know how the DIY story goes, because the gist itself warned you: the wiki is a weekend, and the maintenance is forever.
Free plan includes 50 articles. Markdown export anytime.
Side by side
| Dimension | OzBrain | DIY LLM wiki |
|---|---|---|
| CostThe honest row: DIY wins on price and always will. | Free 50 articles, then by size | Free forever |
| The librarianRouting, contradiction checks, and pruning: the weekly work that kills wikis. | Designed into the product | You, every week |
| Contradictions | Caught on write | Found when they mislead |
| ReachA shared brain reaches every agent you use and every person you grant. | Every agent, plus people you grant | One machine, wired per agent |
| Ownership | Markdown export anytime | Files on your disk |
The wiki was never the hard part
Karpathy named the failure mode in the gist: humans abandon wikis because the maintenance burden grows faster than the value. Routing new facts to the right page, noticing when two pages disagree, pruning what stopped being true. That work lands on you every week for as long as the wiki lives, and the weeks you skip are the weeks it rots.
OzBrain hosts the librarian work as product design. New knowledge lands in staging and gets routed to the article it belongs to. A write that contradicts what the brain holds pauses, and the contradiction surfaces instead of silently overwriting. Continuous maintenance is the intended behavior: write-time size discipline, a knowledge split so agents read only what they need, and every version records which agent wrote it and when.
Reach is the other half, and it has two parts. A wiki on your laptop feeds the agents on your laptop, after you wire each one; the brain reaches Claude, ChatGPT, Claude Code, and Cursor with one URL and one approval. And a DIY wiki has exactly one librarian and one reader for life: you. A brain takes a grant, and a collaborator’s agents read the same articles yours do, with every write attributed.
The pattern you already run, with the librarian work off your plate.
Host my wikiWhat building it yourself bought you
The files are yours, the price is zero, and the structure bends any way you want on any morning. Building one also teaches the pattern better than any landing page: nobody who has run an LLM wiki goes back to believing platform memory is enough.
OzBrain keeps the part of DIY worth keeping. Markdown in, markdown out, the whole brain exportable anytime.
Import the wiki you already built
Connect OzBrain to your agent and ask it to stage your pages as articles. Each one passes through routing and contradiction checks on the way in, which doubles as the cleanup pass your wiki has been owed for months: the duplicates, the stale facts, and the pages that quietly disagree all surface as they land.
If you ever want the files back on your disk, the export is one click and the format is the one you started with.
Keep the pattern. Retire from the librarian job.
Everything the gist promised still holds: agents that build and maintain the knowledge base instead of you. What changes is that continuous maintenance is designed into the product rather than left as your weekly chore, the same wiki answers in every agent you use, and the collaborators you grant access get it behind their agents too.
Import the wiki tonight; write-time size discipline is live, and continuous maintenance is the designed behavior. Sharing ships with Pro and Max.
Common questions
What did Karpathy actually publish?
A gist describing a knowledge base your AI agent builds and maintains instead of you, with the warning that hand-run wikis die of maintenance. It collected 5,000 stars on GitHub in three months. The link is on our homepage; the pattern is exactly what OzBrain hosts.
Is OzBrain affiliated with Karpathy?
No. The gist is public and the pattern belongs to nobody. We built the hosted version and cite him because the source matters.
My wiki is in git. Do I lose the history?
Your git history stays yours, and the brain starts its own: every article version is kept with its author and timestamp, across every agent that writes. Version history is native rather than bolted on.
Can I share the brain with a collaborator?
Yes, and this is the part DIY cannot reach: grant their account read or read and write access, and their agents read the same brain yours do. Two writers on one article resolve through the staged write protocol. Sharing ships with Pro and Max.