The model isn’t the memory

In April, I was asked what would break if the company behind my AI changed its pricing. My answer was that the conversational layer would break. The notes underneath it should still be there.

By August I was moving my work to a different model. That turned a tidy answer into a practical test.

The part I wanted to keep wasn’t the chat window. It was the accumulated context: what I was building, what I’d already decided, how I wanted the work done, the mistakes I didn’t want repeated. A new model with none of that would be a clever stranger. I’d be starting the working relationship again.

So I’d been putting that context somewhere I could carry it. Plain notes. Instructions. Skills. Project files. Things a different tool could read, rather than things I hoped one product would remember about me.

It didn’t make the change effortless. Models behave differently. A rule that one follows can need explaining again to another. Connections break. You find out which parts of the setup were less portable than you thought. Keeping the files gives you a starting point; it doesn’t do the move for you.

Still, there’s a useful difference between changing the reader and rebuilding the record.

I want to be able to choose a model because it’s good at today’s work. I don’t want years of accumulated context making that choice for me. If all my useful history stays inside one service, leaving has a cost that grows every time I use it. The better the relationship gets, the harder it becomes to replace.

That’s an odd thing to call loyalty. Sometimes it’s just moving being a pain.

This is the part of AI ownership I find most practical. I can use a hosted model and still hold the material that makes it useful to me. I can rent the capability without letting the entire working relationship become something I can’t take elsewhere.

There are limits. Local files don’t make a conversation private if I send them to a hosted service. Owning the notes, running the model locally and controlling what leaves the machine are different questions. I need to answer each honestly. A folder on my laptop doesn’t settle all three.

Nor does a large archive guarantee useful memory. The right note still has to be found. A stale decision still has to be corrected. A confident summary can preserve the wrong thing beautifully. I care about whether the system can return to the source, rather than how convincingly it tells me it remembers.

The small test I’d use is this: take one live project and imagine opening it with a new model tomorrow. Could you hand over the goal, the important decisions, the current state and the instructions without spending the morning reconstructing them?

If you can, you’ve kept something valuable outside the tool. If you can’t, write those things down before adding another conversation to the pile.

I’ll keep changing the tools. I want the work to come with me.