There's a feature in modern software that nobody puts on the box: once you're in, leaving hurts.
It used to be called vendor lock-in, and it used to be a complaint. In consumer AI, it's been rebranded as "memory," and it's sold as a benefit. The longer you use a tool, the better it knows you, the less you want to start over. That's not a side effect. That's the strategy.
Memory as a retention mechanism
Think about what happens when you turn on memory in a chat assistant. Over months, you teach it your context: your work, your preferences, your recurring problems, the shape of your life. The assistant gets more useful. You get more attached. The cost of switching — to a different model, a different company, a different stack — climbs every week.
This is, from a product perspective, brilliant. From a user perspective, it's a trap with a friendly face. The thing you've built is not portable. It's not yours. It's a profile of you, held by someone else, that stops working the moment you stop paying or stop using their product.
Lock-in used to be something companies did to you with file formats and APIs. Now it's something you do to yourself, one conversation at a time, and you thank them for it.
The asymmetry of memory
Here's the part that's hard to see until you try to leave: you and the company are not holding the same thing.
You hold: a vague sense that the assistant "knows you." A feeling of convenience. Maybe, if you're careful, an export of your chat history.
They hold: a structured, queryable, proprietary representation of your patterns, preferences, and intentions. The thing that actually makes the assistant useful. The thing you cannot take with you.
You're not equal parties to this memory. You contribute. They own. When the relationship ends, you walk away with the raw text. They keep the model.
Why this is worse than old-school lock-in
Old lock-in was annoying but recoverable. You could migrate a database, rewrite integrations, convert files. Painful, but possible. The new lock-in is built on something you can't export: a neural network's internal representation of who you are. There is no "export my personality profile" button. There is no standard format for "what this model has inferred about me." You can't port it because it's not a thing they'll give you, and it's not a thing you can reconstruct.
This means the switching cost is no longer a technical problem. It's a relationship problem. You don't migrate. You start over. And starting over means months of re-explaining yourself to a new assistant that doesn't know you yet.
The honest version of the deal
If companies were upfront about it, the marketing would say: "Use our memory feature, and we'll make it expensive for you to ever leave." They don't say that. They say "personalized," "seamless," "remembers what matters." All true. All of it in service of the same goal.
There's nothing wrong with a product trying to keep you. There's something wrong with the keeping being structural — with the thing that makes you stay being something you can't take with you, that you built with your own words and time, that you have no claim to.
What it looks like to fix the architecture
The fix isn't a better export policy. The fix is a different place for the memory to live.
Eigen Mesh is built on a simple inversion: your memory is not a feature of an AI tool. Your AI tools are clients of your memory. The memory is yours, separate, portable, and available to whatever assistant you happen to be using today.
You capture what you think and plan — in your words, when you mean it. Eigen Mesh organizes it and makes it available to any AI tool that wants to ask. Switch from one assistant to another next week, and your memory comes with you. Not because a company is generous about exports, but because the memory was never theirs to begin with.
This is the difference between "they let me leave" and "I don't have to ask."
The thing about forgetting
The title of this post is a little unfair. The AI doesn't forget. It remembers perfectly, for as long as they keep it. What it does is hold the memory hostage — and call it a feature.
Real forgetting, the kind you control, is part of owning a memory. You decide what to keep, what to revise, what to let go. With a company-held memory, you decide none of that in any meaningful way. You get a settings panel. They get the architecture.
If you're going to spend the next year teaching an assistant who you are, teach one you can take with you.
Eigen Mesh is open source under Apache 2.0 — github.com/hikaru90/eigen. Self-host with Docker Compose, or get early access to the hosted version. 100 free credits on signup, then €3.99/month. Servers in the EU, and we don't train on your data.