Google Described the Format Hjarni Already Speaks

Google Cloud just published a spec called the Open Knowledge Format. It is a way to write down what an organization knows so that AI agents can read it.

The format is a folder of Markdown files. Each file has a little block of structured data at the top. Files link to each other. A folder can include an index for navigation and a log for history. The whole thing is meant to be read by both people and AI.

We read it and laughed. That is Hjarni. We have been building it for a year.

So we shipped a button. You can now export any folder, any team space, or your entire knowledge base as an Open Knowledge Format bundle.

What the Open Knowledge Format actually is

Strip away the announcement and OKF is refreshingly small. Version 0.1 fits on one page.

A knowledge base is a folder tree. Every concept is one Markdown file. At the top of each file is YAML frontmatter: a type, a title, a description, a link to the source, some tags, a timestamp. Only type is required. Everything else is optional.

Files reference each other with plain Markdown links, which turns the folder into a graph. Each folder can have an index.md that acts as a map. A log.md records what changed.

That is the entire format. No database. No SDK. No proprietary API. No account to sign into. A producer can generate OKF from anything. A consumer can read it with anything. The files live in version control and survive being moved between tools.

Google's reason for building it: knowledge inside most companies is scattered across catalogs, wikis, shared drives, and the heads of a few engineers. An AI agent cannot use any of that until someone assembles it into context the model can see. OKF is a common shape for that context, so every team does not have to invent its own.

It is a good idea. It is also, almost field for field, the thing Hjarni already does.

Why a format matters now

The interesting part is not the file layout. It is the admission underneath it.

For years the gold standard for a note was "human-readable." Clear enough that a person could open it and understand it. That is no longer enough. Knowledge now has to be legible to the agents working alongside you too. We wrote about this shift in agent-legible knowledge: capture was never the hard part, and it definitely is not the hard part now. The hard part is getting the right context back out when you, a teammate, or an AI needs it.

An assistant can only use what it can find and interpret. If your knowledge is trapped in a tool that only speaks its own format, most of it does not exist at the moment it is needed. Generation is solved. Finding it again is the problem.

A shared, open format is how knowledge stops being trapped. Plain Markdown with a bit of structure is a better substrate for this than a proprietary store. Google picking the same substrate is a strong signal that the rest of the industry is moving this way too.

Hjarni already speaks it

Here is the mapping, because it is almost funny.

OKF wants Markdown files with frontmatter. Hjarni notes are Markdown, and every note carries a title, a summary, a source link, tags, and timestamps.

OKF wants a folder tree. Hjarni has folders, nested as deep as you like.

OKF wants files that link to each other. Hjarni has wiki-links between notes, which form exactly that graph.

OKF allows an index.md per folder. Hjarni already recognizes hub notes that mostly list links to other notes.

OKF wants the bundle to be readable by agents. Hjarni has a built-in MCP server, so Claude and ChatGPT already read your notes, follow your instructions, and write notes back in your structure.

The building blocks are the same: notes, folders, tags. We did not adopt OKF so much as discover we had been producing it in all but name.

Export is ownership

The export button is the point we actually care about.

Your knowledge base should not be a hole you pour things into. It should be yours to take. So an Open Knowledge Format export is a clean, self-contained bundle. Your notes as Markdown. Your folders as directories. Your wiki-links rewritten as relative Markdown links so the graph still works on disk. An index for every folder, and a change log. Unzip it and it makes sense without Hjarni anywhere in the picture.

Put it in a Git repo. Hand it to a different tool. Feed it to any agent that reads OKF. Read it yourself in five years. The format is the contract, and the contract is open. That is what we mean by export and ownership: not a panic button you press when you leave, but proof that the knowledge was always yours.

This is the same reason we bet on Markdown in the first place. The best note is one that works twice. Once for you, once for your AI. A note that is also a portable, agent-readable document works a third time: anywhere else you want to take it.

Build a brain, not a document dump

A format is only half the job. A folder of perfectly formatted files is still useless if it is a flat pile.

The value comes from shape. Summaries that compress a note into a handle. Links that connect a decision to the thing that caused it. Folders with a clear purpose. Instructions that tell an assistant how to treat one area of your knowledge differently from another. This is the LLM wiki idea: a brain the agent helps maintain, not a document you keep re-pasting into a chat window.

OKF gives that brain a passport. It can travel in a format that other tools, including Google's own Knowledge Catalog, know how to read. The structure you build inside Hjarni is the structure that comes out.

Try it

Open any folder and click Export. Pick Open Knowledge Format. You get a ZIP that any person or any OKF consumer can read.

If you do not have anything to export yet, that is the better place to start. Make one folder, give it a purpose, and write three notes with real summaries. Connect Claude or ChatGPT over MCP and ask what you already know about that topic. The structure you build is agent-legible from day one, and now it is portable too.

Hjarni is free to start, no credit card. Google just told the rest of the industry which way to point. We were already there.