We didn't build Eigen Mesh because we saw a market opportunity. We built it because we got tired of asking permission.
Permission to use our own context. Permission to take it with us when we switched tools. Permission to know what was being done with the things we had told an assistant in confidence. Permission to trust that the policy we agreed to today would still be the policy tomorrow. We kept waiting for one of the labs to ship the version of memory we actually wanted — owned, portable, honest — and they kept shipping the version that served them instead. So we built it ourselves.
This is a founder's note, not a product brief. The product is at github.com/hikaru90/eigen. What follows is why it exists.
The thing that finally broke us
It wasn't a scandal. There was no leak, no breach, no dramatic moment where trust collapsed. It was slower and duller than that.
It was the day we realized that a year of context we had carefully built up in one assistant was not portable. We had told it about our projects, our preferences, our recurring questions, the way we liked things phrased. We had, in the language of the industry, "trained it on ourselves." And then we wanted to try a different model for a specific task, and we discovered there was no export. There was no API to pull the memory. There was no way to hand it to the other tool. The year of context was theirs, not ours, and they had never intended it to be otherwise.
That is the moment most people shrug and move on. We didn't. We started asking what it would take to build the version where the memory was ours, and the assistants came to it.
The first answer was wrong
The first answer we reached for was the obvious one: build it on top of one of the existing memory kits. There are good ones. mem0, OpenBrain, the various RAG-on-a-laptop setups. They are real, they work, and the people building them are doing genuine work.
We used them. We stopped. The reason we stopped was not that they were bad. It was that they were kits. They assumed we wanted to assemble a stack — pick an embedding model, pick a vector store, wire up a retrieval pipeline, write the glue. That is a fun weekend if you are the kind of person who enjoys that weekend, and we are. But it is not a product. It is not the thing you would recommend to a friend who is not you. It is not the thing that scales to people who want memory they own without also wanting a homelab.
We realized that the gap in the market was not "another framework." It was "a product, that you can also run yourself, that does not require you to be the kind of person who runs things yourself." Open source that only serves the people who would have built it anyway is not really open source. It is a hobby with a license.
So we built the product version. Apache 2.0, self-hostable, but also: install it and it works. Use the hosted version and it works. The self-hosted version is not a consolation prize for people who can't afford the hosted one. It is the guarantee that the hosted one can be trusted, because you can always leave with the whole thing intact.
The decisions we made on purpose
A few decisions shaped Eigen Mesh, and they are worth naming because they are the opposite of what the labs would do.
We don't train on your data. Not because we are nice — because the second we did, we would become the thing we built this to escape. The architecture doesn't require it, and the business model doesn't depend on it. There is no other revenue stream hiding behind the price. You pay us, or you run it yourself, and that is the whole relationship.
We don't infer what you care about by mining your inbox, your transcripts, your PDFs. The big AI memory products do this, and it is the source of most of their value and most of their creepiness. Eigen Mesh works the other way: you say it, you meant it, it's saved. Deliberate capture instead of passive surveillance. The dataset is smaller, sharper, and yours. You chose every line of it.
We don't lock the memory to one assistant. The whole point is that the memory is portable. Whatever you are using this month — ChatGPT, Claude, a local model, something that doesn't exist yet — comes to your memory to get context. Your memory doesn't get uploaded into theirs. The day you switch tools, your context switches with you.
And we made it EU-based, GDPR-shaped, and self-hostable, because the trust question deserves a real answer, not a paragraph in a privacy policy. The strongest version of "you can trust us" is "you don't have to."
What we are not doing
We are not building a chatbot. There are enough chatbots. We are not building a model. There are enough models. We are not trying to win the assistant race. We are trying to own one layer of the stack — the layer where your data lives — and make it the one layer that is yours.
We are also not pretending open source is a magic wand. Code you can read is only useful if you read it. Self-hosting is only a real option if it actually works. We are aware that the gap between "Apache 2.0" and "a person can actually run this" is where most open source projects die, and we are building to close that gap, not to wave at it.
The bet we are making
The bet is simple. We are betting that enough people are tired of asking permission that a memory they own is worth €3.99 a month to them, or worth the effort of running it themselves. We are betting that the labs will not, structurally cannot, build this — because their entire business depends on the memory living with them. We are betting that open source, done as a product and not a hobby, can take the one layer that matters most.
So far the bet is holding. People are tired of asking permission. We were too. That is why we are here.
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, then €3.99/month — less than a coffee. EU servers, GDPR, no training on your data. Your memory. Not theirs.