The memory dies when the model does

I spent months teaching an assistant how I think. It learned my projects, my habits, the way I circle a problem for a while before I commit to it. That’s real work, done a little at a time, in a hundred small corrections. It was also the most valuable thing in the whole arrangement.

Then I tried to take it somewhere else.

The thing nobody mentions when you sign up

Every assistant that remembers you keeps that memory inside its own walls. The model and the memory ship together. It feels normal, even convenient, right up until the day you want to leave.

A better model shows up from someone else. Or the company changes its terms. Or you just want a second assistant’s read on the same problem, for a different angle. That is when you find the months of teaching don’t come with you. They were never yours to carry. They were a feature of the room you were standing in, and you can’t take the room.

There’s a name for this everywhere else in software: lock-in. The cost of leaving, raised high enough that you stay. I usually spot it in file formats and contracts. Here it’s quieter and far more personal, because the thing holding you isn’t a format. It’s the months of yourself you poured in. A memory that only works inside one company’s model is the most effective retention tool ever built, and it never looks like one. It looks like the assistant simply knowing you.

So you start over. New assistant, blank slate, explaining yourself from scratch as if the last half-year never happened. Most people don’t even count it as a loss. They sigh, re-type their life, and assume that’s how it has to be.

Memory shouldn’t live inside the model

Here’s the assumption worth questioning: that the memory and the model are one thing.

They aren’t, and you can feel the seam. The model is the part that reasons, fast and fluent, replaceable the moment a better one arrives. The memory is the part that knows you, slow to build and genuinely yours. They got bundled because that’s how the first products happened to be built, not because they belong together. And the bundle has stayed welded shut for a reason: it works beautifully for the people selling the model. The harder it is to leave with your memory, the less likely you are to leave at all. No incumbent wakes up wanting to fix that. One is a tool you should swap freely. The other is a record you should never lose, and right now it’s being used to hold you in place.

Pull them apart and the whole picture changes. Let the memory sit on its own, outside any single model, and the model becomes what it always should have been: something you can switch out. When a sharper one ships, you point it at the same memory and it picks up where the last one stopped. Nothing to re-teach. The knowing stays put while the reasoning gets better.

More than one at a time

Once the memory lives on its own, something stranger opens up. Not one model at a time, but several at once, reading the same record.

Ask Claude something this morning, and what it works out is waiting for a different assistant this afternoon, and for the small private model running your own app tonight. These aren’t three copies drifting out of sync, but one shared store with several instruments taking turns, each adding to what the others can draw on. What one of them works out becomes ground the next one stands on.

I’ll admit that the first time this clicked, I laughed. Everyone had told me the memory belongs to the model, and it simply doesn’t. That was the convenient story, not the necessary one.

What you’re actually protecting

Which assistant is best this quarter is the easy question. They’ll keep leapfrogging each other, and you should keep switching to whichever is sharpest. That part is easy to replace.

The hard part, the part worth guarding, is everything they learned about you along the way. Hold that outside the model and the leapfrogging costs you nothing. Hold it inside, and every upgrade is also a small erasure, and every day is one more day you can’t afford to leave.

So the next time something offers to remember you, don’t ask how good its memory is. Ask where it lives, and who holds the key. If it lives inside the model and the key stays with the company, then the better that memory gets, the less free you are to go. The thing learning all about you is also the thing keeping you.