LLM wiki: Obsidian vs Hjarni

Andrej Karpathy's LLM wiki gist describes an Obsidian vault and an agent that maintains it. It is a good setup. For a lot of people it is the right one.

But "Obsidian or Hjarni" is the question we get most from people trying the pattern. Both run the same idea: a Markdown wiki an LLM keeps current. They differ in where the notes live and how many places can reach them.

Here is the honest comparison.

The pattern is identical

Start with what does not change. Both are notes, folders, tags, and links in Markdown. Both have an LLM doing the bookkeeping: filing, tagging, linking, flagging contradictions. Both pay the synthesis tax once at write time instead of re-deriving knowledge on every query.

If you have read why Karpathy's LLM wiki is right, that is the part both options get right. The difference is everything around the notes.

Obsidian: local files you own

An Obsidian LLM wiki is a folder of .md files on your machine, with Claude Code or a similar agent editing them.

What that gives you:

• Local files. Your notes are a folder you can grep, back up, and open in any editor.
• Git history. Branch it, diff it, roll it back. Knowledge with a changelog.
• A graph view. Obsidian's force-directed graph is genuinely nice for seeing how notes connect.
• Plugins. A large community ecosystem, if you like to tinker.

The trade-off is reach. As far as we have seen, the vault lives on one machine and one client edits it well. You are at someone else's place, you remember a thing, and adding it from your phone is awkward. ChatGPT cannot see the vault that Claude Code edits. None of that is a dealbreaker. It is friction, and friction is what kills knowledge habits.

Hjarni: hosted, over MCP

Hjarni is the same wiki pattern hosted, with a built-in MCP server so any MCP-capable client reads and writes the same notes.

What that gives you:

• Every device. Capture on your phone, refine on your laptop, query next week. No syncing.
• Every client. Claude, ChatGPT, and Cursor read the same notes, folders, and tags. Start a note in one, finish it in another.
• Write-back from anywhere. You talk to whatever AI you are already in and it writes to the wiki. You do not open a separate app to add a note.
• Sharing that stays live. A folder you share is a wiki other people query and add to from their own AI, not a static repo.

What you give up with Hjarni

The honest list, the same one we put in the Karpathy post:

• No git history. You can update notes safely, but it is not git log. If you want branchable, diffable knowledge, the local vault wins.
• No graph view. Hjarni shows links between notes. The force-directed graph is an Obsidian thing.
• No local filesystem. Your notes live in a hosted knowledge base, not a folder of files you grep. For some people that is a hard no, and that is fair.
• No plugin ecosystem. You trade a marketplace for a focused product.

If those matter to you, run the Obsidian version. We mean that.

Who should pick which

Obsidian if: you live in a terminal, you want git history and a graph view, you like local files, and "only on this laptop" does not bother you.

Hjarni if: you want the wiki everywhere and in more than one AI, with no syncing, and you would rather not maintain the plumbing.

It is the same pattern either way. The choice is local control versus reach.

You do not have to decide blind

You can try the hosted version without throwing away your vault. Import from Obsidian brings your folders, frontmatter, wiki-links, and attachments across, and you can export back to Markdown whenever you want.

Still weighing it feature by feature? Hjarni vs Obsidian lays out the full comparison. And if you are earlier than this, still wondering whether you need more than a single Markdown file, LLM wiki vs. plain Markdown is the place to start.