Every New Model Starts From Zero. Your Company Doesn't Have To.

Switching AI models shouldn't mean starting from zero. Keep your context in plain files you own, connected by MCP, so any model picks up where you left off.

Has this ever happened to you?

A new model comes out. Everybody says it's way smarter. You open it up, excited, and ask it to help with the proposal for your biggest client.

And it has no idea who that client is.

It doesn't know what you quoted them last time. It doesn't know you stopped using that supplier in March. It doesn't know your pricing rules, your team, or the decision you made two weeks ago that changes everything about this proposal. So you paste in the same three paragraphs of background you pasted into the last model. And the one before that.

I've done this so many times I could recite my own business to a robot in my sleep. I basically onboard a new hire every time a model drops, and the new hire never remembers the orientation.

Here's the fix, in plain words, and you can start it this afternoon.

Short answer: keep what your business knows in plain text files you own, with a short index at the top, and connect them to your AI tools through MCP, a standard plug that Claude, ChatGPT, Gemini and others can use. Then any model reads the same files and starts where the last one stopped. That setup is called a company brain.

Why every model starts from zero

When you use an AI tool, the stuff it "remembers" about you lives inside that tool. Your chat history, your saved memories, the files you uploaded to a project. All of it belongs to that app.

Switch apps, and none of it comes with you. Switch models inside the same app, and sometimes it does and sometimes it doesn't. Either way, the knowledge is sitting in a place you don't control.

There's a second problem people miss. Even when a model can see your history, it has to read it. Models charge by how much text they read and write, counted in tokens. If your history is a long pile of chats, the model re-reads that whole pile to catch up, and you pay for the re-reading every single time.

Atlassian made a point of this in its October 2026 MCP announcement: their rebuilt connector uses up to 25% fewer tokens than the old one in their own test. A company that size doesn't brag about tokens unless tokens cost real money. They do.

And it gets worse with a team. Your sales lead uses one assistant, your office manager uses another, and you use whichever one you opened last. That's three separate memories of the same business, none of them complete, none of them talking. Ask all three the same question about the same client and you can get three different answers. The knowledge exists. It's just scattered across apps that don't share.

So you've got two leaks. The knowledge is trapped in one tool, and the catch up is expensive.

MCP, explained like you're nine

You've got a lamp. You've got a wall outlet. You don't care who made the lamp or who wired the house. You plug it in and it works, because everybody agreed on the shape of the plug.

MCP is that plug, for AI.

It stands for Model Context Protocol. It's an agreed shape that lets an AI model connect to a tool or a folder and read from it. Any model built to use the plug can connect to anything that offers one.

That's why it matters for your business. Atlassian's own MCP setup docs list Claude, ChatGPT, Codex, Cursor, VS Code, Gemini and more as clients that can connect. Different companies, same plug. Atlassian says its server has nearly 2 million monthly active users and handles 15 million plus tool calls a day, per their October 2026 press release. The plug is real and people use it.

Here's the part that makes it yours. If your company knowledge sits behind an MCP connection, you can swap the lamp anytime. Claude today, ChatGPT tomorrow, whatever ships next spring after that. The outlet stays put.

Why plain files outlive any model

Now for the piece I care about most.

The plug only helps if what's behind it belongs to you. If your knowledge lives inside one vendor's app, the plug just gives a new model access to someone else's house.

So put the knowledge in plain text files. Regular files, the kind that open in Notepad, sitting in a folder you control. I walked through the whole layout in the original company brain post, but the short version:

• One page per thing that matters: each client, each supplier, each process, each big decision.
• Pages link to each other, so you can walk from a client to the products they buy to the vendor who makes them.
• A short index at the top that lists every page with one line about it.
• A log that gets one dated line every time something changes.

Text files have been readable for decades and nobody can raise the price on them. Every model coming out over the next few years will be able to read a folder of text. That's about the safest bet in this whole field.

The model becomes the reader. Your files stay the memory. When the reader gets an upgrade, the memory doesn't care. That's the design behind every company brain I build.

Start from a short index, not the whole history

This is the trick that kills the expensive catch up.

Instead of making a model re-read everything you've ever said, you give it a short index first. A page or two that says what exists, what's hot right now, and where to look for details. The model reads that, figures out what it needs, and opens only those pages.

Think of it like walking into a library. You don't read every book to find one fact. You check the catalog, and then you walk to one shelf.

Here's my own setup, because I'd rather show you than describe it. Before my assistant does any work, it reads three short index files at the start of every session. That's about 1 to 2 thousand tokens, per my own operating rules. Then it works. It doesn't re-derive what the record already says.

When a new model comes out, I point it at the same three files and it's caught up before I've finished my coffee. Same folder. Different reader. No orientation speech.

What goes in first

People freeze here because they think they have to write down everything before it's useful. You don't.

Start with the stuff you keep re-explaining. If you've pasted the same background into a chat window more than twice, that's a page. Your biggest clients. How you price a job. Who on your team handles what. The decisions that took a long meeting to reach and that everybody forgets by Thursday.

Then add one thing every week. A call recap. A quote that taught you something. The reason you dropped a vendor. Small and boring beats big and abandoned.

After a month or two, the index starts doing the explaining for you. You'll notice the first time a new model answers a question about your business correctly without you typing a word of background. It's a weirdly good feeling.

What a brain won't do

There's always a boundary, so here it is.

It won't remember what nobody wrote down. A brain is only as good as what goes into it. If a decision got made on a phone call and nobody logged it, no model on earth can find it. The habit of writing things down is the whole engine.

It won't make a weak model smart. A better model reads your brain better. A worse one reads it worse. The brain makes sure every model starts with the same facts, and that's all it promises.

It isn't automatically private. An MCP connection can see whatever you let it see. You decide what goes in the folder and who gets connected. You make that choice on purpose.

The obvious part nobody says out loud

Every AI vendor wants your memory living in their app. Of course they do. It's the reason you can't leave, and they know it damn well.

I'm not mad about it. It's just how the business works. But you don't have to play along. Write it down in files you own, put a plug on the front, and let the vendors compete for the right to read it.

If you want to see what the different ways of doing this cost, including Atlassian's version, I laid them side by side in what a company brain costs in 2026. And if you're weighing Atlassian's new Teamwork Graph specifically, here's what it means if you're not running Jira.

Do this this afternoon

You don't need to buy anything to start.

1. Make a folder called brain.
2. Make one file in it called index.md.
3. Write ten lines in it. Your business in one sentence. Your top five clients or customers, one line each. Your three biggest active projects. One line on where everything else lives.
4. Next time you open any AI tool, paste that file in first, before you ask anything.

Notice how much less you have to explain. That's the feeling of a model starting from your index instead of from zero. Now imagine that index pointing to a hundred clean pages, and the model reading it on its own through a plug.

Want it built for you?

That's the Company Brain Build. Your notes, docs and tools turned into one brain of plain files you own, with an MCP connection so Claude, ChatGPT or whatever's next can read it. It's $3,000 to $10,000, and you keep every file if we ever stop working together.

You can book a brain call or call me at 801-316-8544. I'm in Salt Lake City, and here's what a build looks like if you're local. Bring the paragraph you keep pasting into every new model. We'll start there.