The Moat Is Your Memory

Every generation of software has had a moat, and it is rarely the thing the founders were proudest of. The desktop era's moat was the file format. You did not stay with a word processor because you loved it. You stayed because ten years of documents would open only in it. The mobile era's moat was the social graph. The product could be mediocre, and often was, but your friends were there, and moving meant moving alone.

The moat of this era is being poured right now, and it is memory. Not the chips. The other kind. What the assistant knows about you.

Why memory is the product

The models themselves are converging. Anyone who has used three of them side by side for a month has noticed that the gap between the best and the second best is narrower than the marketing suggests, and it narrows again every quarter. A language model wants nothing and knows nothing about you in particular. On its own it is a very capable stranger, and a stranger, however capable, has to be briefed every time.

The briefing is where the value is. An assistant that has seen a year of your work knows which of your projects is the one you actually care about, which colleague needs the long version and which needs three lines, what you meant the last four times you said "the usual." The first week with a new assistant is mostly you explaining yourself. The fiftieth week is the assistant finishing your sentences. That difference is not the model. It is the accumulated context, and the context does not come with you when you leave.

Which is the whole point. A vendor who holds a year of your context does not need a better model to keep you. They need only to make leaving feel like starting over, and starting over, with a tool you use every hour, is a cost most people will not pay twice.

I have seen this movie before

I spent a long stretch of my career in payments, and payments taught me that the lock-in is never the feature on the brochure. It is the integration nobody wants to redo, the transaction history nobody wants to migrate, the reconciliation logic that grew around one vendor's quirks until it could not be pulled loose. Customers stayed with providers they disliked for a decade because the cost of leaving was measured in institutional memory, and institutional memory is the most expensive thing to rebuild.

Personal memory is about to work the same way, at the scale of every individual who uses these tools, which within a few years is most of the working world. The difference is that when a company is locked in, it has lawyers and leverage. When a person is locked in, they have a settings page.

What is actually being held

It is worth being precise about what the memory contains, because "context" sounds harmless. It is the draft of the email you never sent. The health question you asked at a low moment. The salary negotiation you rehearsed. The name of the person you are quietly worried about. The full map of how you think, assembled from thousands of small requests, each of which felt trivial at the time. No social network ever held anything like it, because a social network only saw what you chose to post. An assistant sees what you were trying to figure out.

Nobody has decided who owns that. The terms of service have opinions. The law, in most places, has not caught up. And the market, left alone, will resolve the question the way markets resolve every lock-in question, which is in favor of whoever holds the asset.

What I would want, as a user and as a builder

I build AI products, so I am arguing partly against my own short-term interest here. I still think the answer is clear.

Memory should be portable. Not in principle. In a file, in a format a competitor can read, exportable in one action, the way you can export your contacts today. If the industry does not build this voluntarily, it will be built by regulation, later, worse, and by people who do not understand the systems.

Memory should be legible. You should be able to read what the assistant believes about you, in plain language, and correct it. Most people would be startled by the picture. Some of it is wrong. Some of it is right in ways they would rather not have written down.

Memory should be deletable, and deletion should mean deletion. Not hidden from the interface. Gone from the weights, gone from the logs, gone from the training set. This is technically harder than it sounds, and vendors who have not built for it will find it very expensive to retrofit, which is exactly why it has to be demanded now rather than later.

The model and the memory should be separable. The deepest version of portability is one where the context lives with the user and any model can be pointed at it. It turns the moat into a commodity, which is why vendors resist it. I think it is where we end up anyway, and the companies that get there first, voluntarily, will be the ones trusted with the most.

The part that is not about business

There is a quieter reason to care. A system that remembers everything about you is not the same as a person who knows you, and the gap between the two is easy to lose sight of when the first one is so convenient. Nobody knows what self-improving systems will be able to do, and I do not pretend to. But a more capable memory is not a more human one. Being known is something that happens between people who can be surprised by each other, who can forget, who can choose not to bring something up. A ledger of everything you ever asked is not that. It is a record. Records are useful. They are also the thing you cannot get back once someone else holds them.

So before you pour another year of yourself into whichever assistant you have settled on, ask the boring question. Can I take this with me. If the answer is no, you are not the customer. You are the moat.