I gave AI a brain, and now it runs half my business.
It took an afternoon to build. It costs almost nothing to run. And it's the single biggest unlock I've found as a founder.
Here's the full system — and exactly how to build it yourself this week.
Most people use AI like a temp worker with amnesia.
Open a chat. Paste some context. Get a response. Close the tab. Next conversation? Start from scratch. Re-explain who you are. Re-explain your business. Hope it gives you something useful.
That's not an AI employee. That's a search engine with a personality.
The problem isn't the AI. It's that you're not giving it anything to remember.
PART ONE: THE KNOWLEDGE BASE
Obsidian is a free note-taking app that stores everything as plain text files on your computer. No subscription. No lock-in. Just folders and markdown files that link to each other.
I turned mine into a full business operating system.
There's a Memory file — think of it as an onboarding doc for an employee who never forgets. It contains who I am, what my business does, how we're structured, our processes, our tools, my communication style, and my goals. Everything Claude needs to know before I say a word.
There's a Client Roster — every active client with their key details, health status, and who's responsible.
There's an Action Tracker — every open task, who owns it, when it's due.
There's a Library of frameworks — sales process, production workflow, org structure. All documented.
There's a Templates folder — for call notes, follow-up emails, proposals, daily briefs. Reusable formats that Claude fills in automatically.
And it all links together. Every file has a parent. Everything connects back to a central Home page. It's a knowledge graph of your business.
You could build the basics of this in an afternoon. The Memory file takes the longest because you're essentially writing a brain dump of everything someone would need to know to do your job alongside you.
But once it exists? Every conversation with AI becomes dramatically more useful. Because it's not starting from zero anymore.
PART TWO: THE AUTOMATIC MEMORY LOOP
This is the part that changed everything for me.
I use Fathom to record and transcribe my calls. Zapier watches my Fathom account and automatically drops every transcript into a folder in Google Drive. No manual step. Every call — client meetings, team standups, discovery calls, one-on-ones — lands in Drive as a full transcript.
Claude Cowork has access to my Google Drive through MCP connectors. So every day, it can pull the latest transcripts and process them.
Here's what "process" actually means:
It reads the raw transcript. Extracts a summary of what was discussed. Pulls out every decision that was made. Identifies every action item — who owns it, what's the deadline. Then it writes all of that to the correct files in my Obsidian vault. Actions go to the Action Tracker. Decisions get logged. Client-specific info gets filed under the right client.
Three weeks ago I was on a call where we agreed to change how we handle product shipping for a specific client. I forgot about it completely. Two days later I asked Claude about that client's status and it pulled up the decision from the transcript — with the exact context of why we made it.
I didn't remember. My system did.
That's the difference between using AI and having an AI employee. An employee remembers what happened in the meeting. Even when you don't.
You don't need Fathom specifically. Any transcription tool that exports to Drive, Dropbox, or a local folder works. Otter, Fireflies, even the built-in recording in Google Meet or Zoom. The point is: get your calls into text, get that text somewhere Claude can read it, and let it do the extraction.
PART THREE: THE INTELLIGENCE LAYER
Obsidian is the brain. But it's a filing cabinet on its own — structured, organised, but passive.
Claude Cowork is what makes it active.
Cowork runs on your desktop and connects to your actual tools through MCP — Model Context Protocol. Think of MCP as giving AI a set of keys to your work environment. Slack, Google Calendar, Gmail, Google Drive, ClickUp — whatever you use. You authorise what it can access and it handles the rest.
This means Claude isn't just reading your vault. It's also reading your Slack channels, checking your calendar, pulling up your Drive files, and cross-referencing everything against the knowledge it already has.
I can say "check my Slack and tell me what's going on across clients" and get a full status report in minutes. Who's on track. Who's blocked. Where feedback is late. What needs my attention. Without opening Slack.
I can say "what do I have coming up this week" and get my calendar pulled alongside relevant context from the vault — which clients I'm meeting, what we discussed last time, what actions are still open.
That's not a chatbot. That's a chief of staff.
THE COMPOUND EFFECT
Here's where this gets interesting.
Every call that gets transcribed and processed adds context to the vault. Every session with Claude ends with a summary written back to the vault. Every decision gets logged. Every action gets tracked.
The vault grows every day. And because Claude reads it at the start of every session, it knows more every time you talk to it.
Week one, it knows the basics — who you are, what you do.
Week four, it knows your clients, your team dynamics, your processes, your communication preferences, and the outcomes of 20 previous conversations.
Week eight, it's catching things you missed. Reminding you of commitments from calls you've forgotten. Flagging overdue actions. Connecting dots across different parts of your business.
It's not getting smarter in the traditional sense. It's getting smarter because the knowledge base it reads keeps growing. Your AI employee onboards itself a little more every day.
HOW TO BUILD THIS YOURSELF
The whole system is five pieces:
1. Obsidian (free) — your structured knowledge base. Set up a vault with a Memory file, a Home page, and folders for whatever you track: clients, calls, actions, templates. Spend an afternoon writing your Memory file. This is the onboarding doc for your AI. Be thorough.
*One tip that saved me a lot of headaches: put your Obsidian vault inside Google Drive. I work across multiple machines — studio desktop, laptop at home, laptop on the road. If your vault lives in a local folder, it's stuck on one machine. Put it in Google Drive (or Dropbox, or iCloud) and it syncs everywhere. Every workstation has the same vault. Every workstation gives Claude the same context. Obsidian supports this natively — just point it at the Drive folder when you create the vault.*
2. Call transcription → Google Drive — any tool that records and transcribes your calls. Route the transcripts to a folder in Drive. Zapier, Make, or even a manual export. The goal is: every call becomes a searchable text file. I use Fathom with a Zapier automation that watches for new transcripts and drops them straight into a "Transcripts" folder in Drive. Zero manual steps. But even a simple workflow where you export and drag the file into a folder works — perfect is the enemy of done here.
3. The Obsidian MCP — this is the bridge. There's an open-source MCP server for Obsidian that gives Claude direct read and write access to your vault. You install it, point it at your vault folder, and suddenly Claude can read every file, create new notes, edit existing ones, move things around, and search across everything. This is what turns Obsidian from a note-taking app into a live system. Without it, you'd be copy-pasting between Claude and your notes like everyone else. With it, Claude just writes straight into your vault. The setup takes about five minutes — it's a Node package you configure in Claude's MCP settings.
4. Claude Cowork + MCP connectors — the intelligence layer. Cowork is Claude running on your desktop with access to your actual tools. You connect each tool through its own MCP connector: the Obsidian MCP for your vault, Google Drive for your transcripts and files, Slack for team communication, Google Calendar for your schedule. Each connector takes a few minutes to set up. The key setting: in your Cowork user preferences, tell Claude to read your Memory file at the start of every session. That single instruction is what gives it persistent context. Without it, you have a powerful tool with no memory. With it, you have an employee.
5. Custom instructions — this is the glue that makes the whole thing work, and it's what makes this portable beyond Cowork. Claude has a "custom instructions" field (in Cowork it's called user preferences, in Claude Projects it's the project instructions). You write one line: "Before answering any question, always search the Obsidian vault for relevant notes. Use what you find to inform your response." That's it. That single instruction means Claude reads your vault before every reply. It works in Cowork, it works in Claude Projects, it works anywhere Claude has access to the MCP. You can also add rules for writing back — tell Claude where to save different types of information (actions to the tracker, decisions to a log, session summaries to a sessions folder). I keep all of this in my Memory file itself, so Claude reads the routing rules at the same time it reads the context. The whole system becomes self-sustaining.
That's it. No code. No complex automations. Obsidian is free. The MCP connectors are free. The only paid pieces are Claude and whatever transcription tool you choose. The actual setup is an afternoon, and the Memory file is most of that time.
I'm not saying this replaces your team. My 15 people are irreplaceable. They make creative decisions, build relationships, and do work that no AI can do.
But the operational overhead? The context switching? The "what did we agree on that call last week?" The scrolling through Slack at 7am trying to figure out what happened overnight?
That stuff doesn't need to be you.
Your tools are only as good as the system behind them.