Last week made something clear that a lot of people in AI have been circling around for months.

On June 9, Anthropic released Fable 5. It was the most powerful model they had ever put in public. State of the art on almost every benchmark. Four days later, on June 12, the United States government issued an export order. Anthropic had to switch Fable 5 off for every foreign national in the world. Inside the country, outside the country, even their own staff who were not US citizens. The reason was a narrow security issue, but the result was simple. A model that thousands of people had started to build on top of was gone in four days.

Not gone because the company changed its mind. Gone because of an order nobody saw coming.

Sit with that for a second. If you had wired your product, your workflow, or your team's daily work to that one model, you woke up on June 13 with a hole where your intelligence used to be.

This is the part most people are missing. The story is not really about one model or one government order. The story is that the model layer is now something that can move under your feet without warning. And if the model can move, then anything valuable you own cannot live inside the model.

Satya Nadella said the same thing, in different words

A day after the Fable news, Satya Nadella published an essay on the future of the firm in an AI economy. People read it as a Microsoft strategy piece. It is more useful than that.

His core point was that the model is not the asset. The asset is the learning loop you build on top of the model. He split it into two kinds of capital. Human capital is the judgment and knowledge of your people. Token capital is the AI capability you build and own. The companies that win, he said, will be the ones that turn their workflows and their corrections and their hard won judgment into a system that gets better every time it is used.

Then he gave the test. And this is the line worth tattooing somewhere.

A company should be able to swap out a general model without losing the company veteran expertise built into their system. That, he said, is the real test of control and sovereignty in this era.

Read that next to the Fable story. The government pulled the general model. The question Nadella is asking is whether you lose your veteran when that happens. If you do, you never owned anything. You were renting intelligence and calling it yours.

So what is the thing you actually own

Here is the honest version. You cannot own a frontier model. That race costs billions and you will lose it. You do not need to.

What you can own is the loop that sits on top of whatever model you happen to be using this month.

Think about how an AI coding agent works today. It reads your code. It makes a change. You correct it. It makes the same mistake next week, because the moment the context window resets, every correction you gave it is gone. The intelligence was real, but it had no memory and no learning. You were the memory. You were the one carrying the lessons from session to session, in your head.

That is the gap. The model is smart and forgetful. The value is in the part that remembers and learns, and right now that part is you, doing it by hand.

What I have been building

I have been building this loop as an actual thing you can install. It is called world-model-mcp, it is open source under MIT, and it is on PyPI.

It is a memory layer that runs alongside your coding agent. It is not tied to one model or one tool. It runs across Claude Code, Cursor, Codex, and others, because it talks over MCP, which is becoming the standard way agents plug into tools.

Here is what it does, in plain terms.

When you correct the agent, it learns the correction as a rule and applies that rule in future sessions, so you stop fixing the same thing twice. It checks proposed changes against what it already knows before the edit lands, so the agent stops inventing functions and APIs that do not exist. When the context window resets, it puts the important rules and recent facts back, so the agent does not lose the thread. It tracks facts over time, so it knows what was true last month and what is true now, and it can tell you when two things it learned contradict each other and pick a winner based on confidence and evidence.

It also forgets on purpose. A correction you gave is worth keeping for a long time. A passing detail from one session should fade. So different kinds of knowledge decay at different rates. A user correction stays strong for two years. A loose session note fades in two weeks. The loop keeps what matters and lets the noise go.

None of this lives in the model. It lives in a local store that you own. Swap the model and the veteran stays. That is the whole point.

Why this matters more after last week, not less
A year ago this would have sounded like a nice optimization. After last week it sounds like insurance.

Models are now things that can be released, jailbroken, and pulled by a government in the same week. Inference is turning into a utility you rent by the token. In that world, the only durable thing is the learning that is yours, sitting somewhere you control, ready to attach to whatever model is standing after the dust settles.

Nadella framed this for big companies and their private knowledge. I am building the version that runs locally, that an individual developer or a small team can own outright, with no cloud holding their data and no single vendor they cannot walk away from. Same idea. The loop is the asset. The model is a part you swap.

The frontier model you depend on today might be gone in four days. The lessons your tools learned from you should not go with it.

If you want to try it

world-model-mcp is on PyPI and the code is open. If you use an AI coding agent, install it, let it watch a few of your sessions, and see whether it stops repeating the mistakes you already corrected once.

And if you build with it or break it, tell me what worked and what did not. The feedback is what shapes what ships next.

The model is temporary. Build the part that lasts.