Your Knowledge Will Outlast the Model
AI providers retired GPT-4 in early 2026. The memory you stored inside it didn't transfer. Here's the question you should have been asking.
In March 2026, both Google and Anthropic launched tools to import your ChatGPT memories into Gemini and Claude. The framing was straightforward: bring your context with you when you switch. The AI memory portability wars had begun.
Except something didn't make it. The explicit facts transferred — name, job title, city, a handful of stated preferences. What didn't transfer was the implicit: the tone calibration built over months of conversation, the task-specific understanding that accumulated across dozens of sessions, the way the model had quietly learned to work with you. Researchers studying the transfers found the resulting context flat, almost biographical — a summary where texture used to be. You got a paragraph about yourself. You left behind something that felt more like a working relationship.
I've been thinking about what that reveals.
The wrong unit of analysis
When people talk about AI memory, they mean something held by the model — a ChatGPT memory list, a Claude project, a Gemini thread. The mental model is that the AI learns about you over time, and that learning is valuable.
That's true, but it's a category error. The AI doesn't know you. It has a record about you, stored in a system you don't control, in a format you can't inspect, attached to a model that will eventually be retired.
GPT-4 launched in March 2023. It was retired in February 2026: 35 months. Microsoft's official policy for its Azure AI services sets model retirement windows at 18 months for GA releases — that's considered generous. The pace has accelerated since: GPT-5.3, 5.4, and 5.5 shipped within a few months of each other in late 2025. The model you're talking to today is already being superseded by something its makers think is better.
Your knowledge has a different lifespan than that.
What survives and what doesn't
A useful thought experiment: if every AI provider shut down tomorrow, what would you have?
If you'd been storing context inside ChatGPT's memory, you'd have a JSON export — if you remembered to request it. If you'd built context inside Claude Projects, you'd have the text of those notes, stripped of any structure the model had learned from them. The implicit part — the preference model, the learned patterns, the calibration — lives in weights and activation states. It's not yours to take.
The March 2026 import tools were a useful stress test of exactly this. What transferred: explicit declarations. The kind of thing you could write in a single paragraph. What didn't transfer: the implicit. How the model had calibrated to your communication style. The context built from long-running projects. The assumptions embedded in hundreds of small corrections.
None of that is transferable because none of it exists as discrete data. It exists as model state — patterns that influenced outputs but were never serialized anywhere you could reach.
Structured knowledge is different. A person record with a name, relationship type, a few notes, and recent interactions — that's a discrete thing you can move. A task linked to a project and a decision — same. A stated preference with a timestamp — same. These don't belong to any model. They work with any model that can read them.
The question people aren't asking
The conversation about AI memory has mostly been framed around capability: which model remembers you best? That's a reasonable question for today. But it's the wrong question for deciding where your knowledge lives.
The right question is: when the model changes — and it will — what survives?
If you're storing context inside the model: your explicit statements, minus everything implicit. A fraction of what you built, delivered as a flat export that any half-decent notes app could have stored for you.
If you're storing context in structured, portable files you control: everything. Immediately usable by whatever model comes next, because it never belonged to the previous one.
This isn't really about format — Markdown, SQLite, plain text. The specific choice matters less than the principle: your knowledge should exist somewhere that doesn't depend on a particular model's continued operation to remain legible. It should be something you could open in a text editor. Something that wouldn't require you to file a data request with a company that may, by then, have been acquired or shut down.
GPT-4 ran for 35 months. Somewhere inside that window, people built up years of context that felt like memory — and then watched it get retired. Not stolen. Not lost in a breach. Just switched off, as planned, because something better came along.
The next cycle is already running.
Asgeir Albretsen is the founder of Harbor.