Video: "Claude Obsidian is INSANE!" by Julian Goldie on YouTube.
The problem: Claude forgets everything
Large language models don't carry memory between sessions by design. Each conversation starts clean. For a one-off question or a quick draft, that's fine. For anything that requires knowing your business — your clients, your code conventions, your tone of voice, what's already been done — it means you're either re-explaining the same context every time, or you're getting generic output that doesn't quite fit.
The obvious workaround is a long system prompt with all your background crammed in. That works up to a point, but there's a token limit, it gets stale, and stuffing a prompt with three pages of business context before every task is wasteful. There's a better way.
What Obsidian brings to the setup
Obsidian is a free, local note-taking application that stores everything as plain markdown files — ordinary text files with standard formatting. That's the key detail. Claude Code can read markdown files directly. Connect the two through the Model Context Protocol (MCP) and Claude can scan your notes before each task, pulling in client context, project briefs, standing instructions, and anything else you've written down.
The result is that Claude's first response on a new task can reference things from your last ten sessions, your client roster, and your brand guidelines — without you typing any of it. It reads your vault the same way a new employee reads a handover document, except it does it in seconds before every piece of work.
How the connection actually works
There are two moving parts: a CLAUDE.md briefing file and the MCP server configuration.
The CLAUDE.md file sits in your project root. Claude reads this at the start of every session — it's your standing brief. You put the project context, naming conventions, what the agent is and isn't allowed to do, and any business-specific facts in here. For most teams, this file becomes the single source of truth for how Claude behaves in this particular project.
The MCP connection gives Claude the ability to read (and optionally write) your Obsidian vault. Julian Goldie walks through two approaches: the plugin route, which is simpler to start but limited to single-file operations, and the Claude Code route, which is more capable — it can reorganise your vault, link notes together, and handle multi-file tasks. For a business that actually wants to build and maintain a useful knowledge base, the Claude Code route is the one worth setting up properly.
The LLM wiki idea behind it
Andrej Karpathy (co-founder of OpenAI) published a note earlier this month about what he called an "LLM wiki" — the idea that you compile raw sources into a structured, AI-readable knowledge base once, and then let the model query it repeatedly rather than re-processing everything from scratch. The Obsidian setup described here is a working implementation of that pattern, available to anyone with a few hours and no budget.
The three slash commands that make it practical day to day: one to ingest a URL and compile it into the wiki, one to process fleeting inbox notes and classify them automatically, and one to audit the wiki for broken links and orphaned pages. Once these are wired up, the knowledge base maintains itself — you add to it naturally, and Claude keeps it organised.
What a small business actually needs to set up
You don't need hundreds of notes to get value from this. A useful starting setup is three folders: one for clients (a note per client — what they do, key contacts, active projects, standing preferences), one for your brand (tone, words to avoid, example copy), and one for project logs (brief notes on what's been done and why). That's it. Probably under 50 notes to begin with, costing nothing to create.
With that in place, Claude produces output that actually sounds like your business, references the right clients by name, and doesn't suggest things you've already ruled out. The difference between a generic AI assistant and one that feels like it knows your business is almost entirely down to the quality of the context you give it — and an Obsidian vault wired through MCP is the cleanest way to do that without rewriting your system prompt every week.
Where this connects to NordSys
Getting Claude Code properly connected to your business context — whether that's an Obsidian vault, a shared Google Drive, your codebase, or a combination — is exactly the kind of setup we handle on our AI Agents care plans. The tools are free. The value is in getting the briefing structure right at the start: which notes to create, how to organise them so Claude can actually use them, and which MCP servers to connect. Most of the gain comes from that initial design, not from the software itself.
See AI agent care plans →