The AI Will Change. Your Record Shouldn't.

A better model arrives every few months, and you will switch to it. What does not survive that switch is everything you taught the old one. That is a design flaw in how the tools are arranged, not an inevitability.

What actually carries over when you switch models?

Make the list honestly. You have spent ninety days with an assistant. What of that is portable?

The conversations, in the sense that you could export them. Not in the sense that the next model can use them — a transcript archive is not working context, it is a pile of scrollback.

The preferences and memories the tool accumulated: gone, or at best re-enterable by hand.

The files: portable, if the tool exports them, and usually as flat current-state copies with no history.

The part that actually took ninety days — the corrections. Every time you said no, we decided the other thing, every constraint you had to restate twice, every nuance that finally landed. That is the expensive asset, and in almost every tool it exists only as a side effect of a conversation that does not travel.

So switching does not cost you a subscription. It costs you the accumulation, and you pay it again from zero.

Day 1, day 30, day 90

The stepper is the argument in three states. Day one is a workspace already in motion. Day thirty is the point where you notice you have stopped repeating yourself. Day ninety reads as a full history — what was decided, what was corrected, who did what.

The critical detail is where that accumulation lives. If it lives inside a model's memory, it is a hostage. If it lives in files you own, signed by whoever wrote each version, it is an asset — and a new model arriving is good news rather than a migration.

Why does the value sit in the record, not the model?

Because of the direction the two are moving.

Models are converging and improving fast, and they are increasingly interchangeable for most work. Any advantage from having picked the right one lasts about a quarter. That is a bad place to keep the thing you cannot afford to lose.

The record moves the other way. It is specific to you, it only accumulates, and it gets more valuable as models improve — a better model reading a richer record produces better work, and adds to the record in turn. Every new model makes a workspace you own more valuable and a workspace you rent more replaceable.

There is a version of this claim that is just vendor lock-in with the sign flipped: stay with us and your context compounds. The honest version has to include the exit. A record you cannot walk out with is not yours, whatever the marketing says. So the test is not whether the accumulation is real. It is whether you can leave with it — every version, every author, intact — and read it somewhere else. If the whole workspace does not export as something plain enough to walk offline, the compounding is happening in someone else's account.

What starting over really costs

Not the setup. Everyone quotes setup and setup is an afternoon.

The cost is the interval where the new tool is confidently wrong in the specific ways the old one had stopped being wrong. It re-suggests the approach you rejected in month one. It uses the term you deprecated. It has no idea which of the two competing documents is the one you actually act on.

You spend weeks re-teaching, and because the teaching is invisible you experience it as the new tool being disappointing rather than as the cost of a migration you did not know you were making.

Do this twice and the pattern becomes obvious: the loop of adopt, accumulate, switch, reset never compounds. It just resets on a cycle set by someone else's release schedule.

The alternative is not loyalty to a model. It is refusing to keep the record inside one.

Where is this argument weakest?

Two places, and they are worth naming rather than waiting to be caught on.

The first: for a lot of AI use, none of this matters. If you use a model to rewrite an email or explain a config file, there is no accumulation to lose and switching costs nothing. The argument only bites where work is produced, kept, and returned to — which is a smaller slice of AI usage than the industry's framing implies, and the slice where the money is.

The second: a record outside every vendor is only as good as what is actually in it. A workspace holding four files and a stale brief does not compound just because its history is signed. The record has to be where the work happens, or it is an archive rather than an asset — which is why the writes have to be first-class and native, not an export you remember to run.

Neither undermines the point. Both change who it applies to, and when.

---

Part four of Inside the workspace. This post supersedes an earlier piece of ours on the same idea, written under an older framing.