The hardest part of Karpathy's LLM wiki pattern isn't the idea.

It's the setup.

A local Obsidian vault. Claude Code in a terminal. A schema file. A log file. A habit you now have to maintain.

Most people won't keep that running.

There's a simpler version: a hosted knowledge base Claude can read, write to, and maintain through MCP, without running anything locally.

What you're building

A structured knowledge base that Claude stays connected to. Not a document you paste in at the start of each conversation. A live system where:

• Claude can fetch your notes when it needs context
• Claude writes new notes back
• Your knowledge compounds across every session

If you want to understand why this pattern works, start here. This post is about how to build it.

Before you start

You need three things:

• A Hjarni account
• Claude Desktop or another MCP-capable Claude client
• Ten minutes to create your first notes

You do not need Obsidian, Claude Code, a terminal setup, or a local folder full of Markdown files.

Step 1: Create your Hjarni account

Go to hjarni.com and create a free account. No credit card required.

Hjarni is a knowledge base with a built-in MCP server. That's what makes the LLM wiki pattern work. Claude connects directly to your notes through MCP instead of reading a static file you paste in.

Step 2: Connect Claude via MCP

Follow the connect Claude via MCP guide. It takes five minutes.

Once connected, Claude can fetch your notes whenever it needs context. You don't paste anything. You don't explain yourself from scratch. Your wiki is available when Claude needs it.

If you use ChatGPT, there's a separate guide for that too. Both clients read from the same knowledge base.

Step 3: Set up your wiki structure

Create a folder structure that matches how you think. A good starter wiki:

/About me
  - Profile
  - Preferences
  - Current goals

/Stack
  - Languages and frameworks
  - Tools
  - Architecture decisions

/Projects
  /Project A
    - Overview
    - Decisions
    - Open questions

/Research
  - Articles and summaries
  - Things to revisit

The Knowledge Wiki template sets this up automatically. Paste the link into Claude and it creates the initial structure, including AI instructions that tell Claude how to maintain your wiki.

Step 4: Write your first note

Start with an "About me" note. Write what you'd normally paste in at the start of a conversation. Be specific.

# About me

I'm a solo developer building an AI-native knowledge base.

## Current projects
- Product development
- Growth and content
- Customer feedback

## Preferences
- Prefer short, direct answers
- Ask before making large architectural changes
- Save useful context back to my wiki

## Current goals
- Publish more useful content
- Improve onboarding
- Make my AI context reusable across sessions

Claude can fetch this note before every answer. You write it once.

Step 5: Let Claude do the bookkeeping

This is where the pattern compounds. Ask Claude to:

• Save a research summary to a specific folder
• Update your stack note when you switch tools
• Create a new note from the conversation you just had
• Link two notes that reference the same topic

Claude writes directly into Hjarni. Next time you open a conversation, on your laptop, your phone, or in a different LLM client, the notes are already there.

Useful prompts

Once Claude is connected, try prompts like:

Read my About me note and use it as context for this conversation.

Save the useful parts of this conversation as a new note in my Research folder.

Update my Stack note with the decision we just made.

Find related notes and link them together.

Create a project overview note from what you know so far.

The point is not to manually maintain the wiki. The point is to make Claude maintain it with you.

What this looks like in practice

You're researching a new approach to something. Claude finds three relevant sources, synthesizes them, and saves a note to your Research folder. A week later, you're in a different conversation. Claude can find and cross-reference that research note when it becomes relevant.

That's the loop. Research in Claude. Save to Hjarni. Act. Repeat.

The wiki grows. The context stays. You stop re-explaining yourself.

The difference from a Markdown file

Karpathy's gist uses a single file you maintain manually. That works until it doesn't. Hjarni gives you:

• Structure: folders, tags, links between notes
• Write-back: Claude updates your notes, not just reads them
• Everywhere: phone, laptop, any MCP-capable client
• Multiple LLMs: Claude and ChatGPT share the same brain

The pattern is the same. The friction is gone.

You don't need to rebuild your context every time you open a chat.

Build the wiki once. Let Claude keep it up to date. Let the context compound.

Set up your LLM wiki in five minutes.

FAQ

Can I build an LLM wiki without Claude Code?

Yes. Hjarni connects through MCP, so any MCP-capable client works. Claude Desktop, ChatGPT, Cursor, or any other client that supports MCP can read and write to the same wiki. No terminal required.

How is this different from pasting notes into Claude?

Pasting is one-shot. An LLM wiki is persistent. Claude fetches what it needs, writes new notes back, and links them. Your context compounds across sessions instead of resetting every time.

Do I need Obsidian to follow this guide?

No. This guide uses Hjarni, a hosted knowledge base with a built-in MCP server. You don't need Obsidian, a local folder, or any file syncing.

Can I use both Claude and ChatGPT with the same wiki?

Yes. Both connect to Hjarni through MCP and share the same notes, tags, and folders. Write a note in Claude, read it in ChatGPT. No syncing needed.