I used ChatGPT's memory feature for about nine months. I let it remember my projects, my writing voice, the names of people I work with. It got good. Scarily good. I'd open a chat and it already knew what I meant by "the Q3 thing" and why I was frustrated about it.
Then I tried to leave, and I understood what I'd actually been doing.
The day I tried to switch
I wanted to try a different model for a week. Just an experiment. I opened the new tool, started a chat, and asked it to help me draft something. It didn't know me. It didn't know my projects, my tone, my recurring questions, my deadlines. I spent the first hour re-explaining things I'd already explained a hundred times to the other one.
That's when it landed: I hadn't been training an assistant. I'd been training a dependency. The memory wasn't a feature I owned. It was a leash. The longer I used it, the harder it was to leave, and the leaving cost was paid entirely by me.
I could export my chats. I couldn't export the model's understanding of me. That lived on their servers, in their weights, in their proprietary representation. I'd spent nine months building equity in someone else's product.
What I actually wanted
I didn't want a smarter ChatGPT. I wanted a memory that knew me, that I could point any tool at. A second brain that wasn't welded to one vendor. If I switched assistants next month, the memory should come with me. If I switched next year, same. If the company shut down, same.
The thing I wanted didn't exist as a product. So I started building it. That became Eigen Mesh.
The shift in mental model
The big change wasn't technical. It was conceptual. I stopped thinking of memory as a feature of an AI tool and started thinking of it as a separate thing I own, like my notes app or my calendar. The assistant is a client. The memory is the server. I am the only person in the middle.
Concretely: when I have a thought — a project idea, a decision, a thing to follow up on — I drop it into Eigen Mesh. It gets organized automatically. When I open whatever AI tool I'm using that day, that context is available to it. I'm not re-explaining. I'm not pasting context. The assistant shows up already caught up.
The first week felt weird. I kept wanting to "tell ChatGPT" things. Then I realized I could just tell my memory, and let the assistant pull from it. The muscle memory changed in about a month.
What I gave up, honestly
A few things. The built-in memory is more tightly integrated — it knows which conversation it's in, it can surface things inline. Eigen Mesh is a separate step. You have to actually capture the thought. That's friction.
But the friction is the point. The built-in memory is frictionless because it's watching everything. Eigen Mesh is a deliberate act: you say it, you meant it, it's saved. The cost of that is one extra step. The benefit is that what gets saved is what you chose, not what they inferred.
I also gave up a little magic. The built-in memory can be spooky in a way Eigen Mesh isn't, because Eigen Mesh only knows what I told it. I'm fine with that. I don't want my AI to be a mind reader. I want it to be a good listener with a perfect memory.
Why I'm not going back
Three reasons. First, I can switch tools now without losing anything. That alone is worth the friction. Second, I can run it myself — Docker Compose, my server, my rules. If our hosted version disappeared tomorrow, my memory wouldn't. Third, and this is the one I didn't expect: I trust my own memory more. I know what's in it because I put it there. I'm not wondering what the model "picked up" about me from three months of chats.
There's a quieter benefit too. Capturing a thought deliberately, in my own words, is itself a kind of thinking. The frictionless version skips that. The frictional version forces me to say what I actually mean. My notes got better. My thinking got a little better.
The version I'd want you to have
You don't have to build this. You can use ours. 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 when you sign up, then €3.99/month — less than a coffee. Servers in the EU. We don't train on your data, because the whole point is that we can't.
If you've ever tried to switch AI tools and felt the cost of starting over, you already understand the pitch. Your memory shouldn't be a hostage.