I’ve been working on this problem for almost a decade.
Not this essay specifically, the problem itself. The question of what it actually means to own your data. What it would take to build a world where the information you generate, the preferences you’ve cultivated, the intellectual fingerprint you leave across thousands of interactions actually belongs to you.
Not in the terms-of-service sense.
Not in the “we respect your privacy” sense.
In the fundamental, structural, architectural sense.
In 2018 I invented UDIF — the Universal Data Interchange Format — because I believed the answer required a new kind of file. A standard that could hold your data in a form that no platform controlled, that any system could read, that lived with you rather than inside someone else’s infrastructure. I built it at Streamlytics, where we processed over ten billion data points from fifty thousand users across a hundred and twenty-five countries. I filed a patent on the underlying system. I watched the data economy evolve around it.
Today, Anthropic launched claude.com/import-memory. Google is testing a similar feature in Gemini. Both are being called portability. But the truth is neither of them actually are.
I know, I know. This can be frustrating but in your rush to move your data from ChatGPT (or any other platform) into Claude you need to understand this:
What Anthropic built is a migration tool.
You paste a prompt into ChatGPT, it generates a summary of what it knows about you, you paste that into Claude’s memory settings. Clean. Simple. Elegant, even. And when it’s done, your memories live in Claude’s infrastructure, encrypted by Anthropic, governed by Anthropic’s terms, subject to whatever policy changes Anthropic decides to make next Tuesday.
You see where I’m going with this…
You basically just changed landlords. You didn’t buy the building.
Google’s version is more transparent about what’s actually happening. When you import your chat history into Gemini, the data is saved to Gemini Activity and used to train Google’s models. They said it out loud. Your years of intellectual output: your questions, your thinking, your creative work, your problems become training data for the next version of a model you don’t own. At least they told you.
I’ve watched this pattern play out across every evolution of the data economy. The platforms that extract value from human behavior are extraordinarily skilled at making extraction feel like empowerment.
First it was your streaming data. Then your social graph. Then your location. Now it’s your cognition; the accumulated record of how you think, what you know, what you care about.
Each time, the same architecture. Your data lives on their servers. Their format. Their rules. Their unilateral right to change those rules without your consent.
The context layer is the new lock-in.
The company that holds your memory holds your switching cost. Every AI interaction you’ve had, every preference you’ve trained into a system, every problem you’ve worked through in conversation; that is an accumulation of intellectual value that compounds over time. And, the platforms know this.
The features they’re launching today aren’t about your freedom. They’re about making sure that accumulated value transfers to them rather than evaporates when you consider leaving.
Real portability looks different.
It means a file that lives on your device or even your own personal cloud. A standard that any AI can read without anyone’s permission. A provenance layer that proves this context is yours; that it came from your mind, your conversations, your creative output and hasn’t been altered. It means the format itself is open, not proprietary.
It means the infrastructure is yours to choose.
I know exactly what this looks like because I designed it. UDIF 2.0 extends the original standard I built for streaming and social data into the AI interaction layer — your conversations, your preferences, your intellectual history, your values, your way of thinking. The Heirloom protocol adds cryptographic provenance: a way to prove that your data is authentic, that it originated from you, that it hasn’t been touched. Not locked inside any company’s system. Yours. On your terms.
This isn’t hypothetical. It isn’t a whitepaper.
It exists because I’ve spent nearly a decade believing that the question of who owns the context layer is the most important question in the data economy — and building the answer.
What Anthropic and Google launched today is a proof of concept for the problem, dressed up as a solution. They’ve confirmed that users want portability. They’ve confirmed that context lock-in is a real switching cost worth addressing. They’ve confirmed that the person who defines the standard for how AI context moves between systems will have shaped something significant.
They just assumed that person would be them.
It isn’t.
The actual alternative isn’t a better migration tool. It’s your own infrastructure. A personal website. Your own data environment where your AI context, your preferences, your intellectual history lives. Not on Claude’s servers, not on Google’s servers; on infrastructure you control. Where you decide what gets connected, what stays private, and which AI systems get access to which parts of your data.
Claude today, something better tomorrow, a specialized agent next year.
Your data doesn’t move. The connections do.
This is what I’ve been building toward with UDIF. The original standard came out of Streamlytics, where we processed billions of data points and I watched firsthand what happens when user data lives inside platform infrastructure with no path out. UDIF 2.0 extends that work into the AI layer; an open format for your AI context that any system can read, that you can host wherever you choose, with cryptographic provenance that proves it originated from you.
Heirloom is the protocol. But the first step isn’t actually a protocol. It’s a mindset shift about where your data should live and who it belongs to.
That's what my first working session was about: How to Let AI Know You Without Letting It Steal Your Work. How to think about organizing your data across your own systems; not just downloading it from one platform and importing it into another. The replay is available if you want to start there.
Otherwise downloading your data from ChatGPT and importing it into Claude is just recreating the same problem on a different platform. Different landlord. Same lease.
Watch the working session replay
Learn more about UDIF
Angela Benton is the inventor of UDIF (Universal Data Interchange Format) and founder of Heirloom, a data sovereignty protocol. She previously founded NewME, the first accelerator for underrepresented founders in Silicon Valley, and Streamlytics, where she built one of the first consumer data portability platforms at scale.
This makes my brain sweat™.
I follow the data sovereignty question from a governance angle rather than a technical one, and the landlord metaphor earns its place and the lock-in argument is sound. Two things gave me pause though: Claude’s import memory feature was live a few days before this published, which slightly changes the 'today' framing. Also worth pushing on the Google transparency point. Telling users their data will train your models, while offering no clear opt-out, is disclosure without recourse. That is a different thing from transparency.