Memory Is the New Lock-In: Why Your AI Profile Should Be Yours

Why do power users end up with data silos?

AI tools change so fast that anyone curious ends up trying several. New ones keep coming, like Muse, Grok Bot and open-source projects. I try them too.

The hard part is always getting a new tool to where the old one was. I've gone back and forth between ChatGPT and Claude. For design work, Cursor worked better for me than Claude, so I moved that part over. Even handing over part of a task was hard. Each tool knew something different about me. None of them knew everything.

At some point I noticed that people like me don't really switch. We use several tools at once and pick one for each job. So we don't just get locked in. Our context gets split up, and every copy is a bit out of date.

What happens to your context when you move between AIs?

Every big assistant now builds a picture of you. OpenAI says ChatGPT's memory can draw on past chats, saved memories, custom instructions, files and connected apps such as Gmail, so you can "spend less time repeating yourself." xAI added a memory feature to Grok in April 2025 that lets it remember details from past conversations. Claude has memory too, and Anthropic offers a built-in flow to import memory from other AI providers.

I find that import tool telling. Nobody builds one unless moving is painful. Here is how it works. Claude gives you a prompt to paste into your old assistant. It begins, "I'm moving to another service and need to export my data. List every memory you have stored about me…" You copy the answer and paste it back into Claude.

So today, to move the most personal data you've ever given software, you ask one chatbot to list what it knows and hope it didn't miss anything. Anthropic is open about the limits. Its help center calls memory imports "experimental" and says Claude "may not always successfully incorporate imported memories."

I don't think anyone is acting in bad faith here. Each vendor keeps your context in its own product and its own format. If you use three assistants, you go through this three times.

Why is memory becoming the new lock-in?

Lock-in used to be about files, contacts or your friends list. With AI assistants, what matters most is context. That means your role, your projects, how you like answers and what you care about. It builds up over time. The longer you use one assistant, the better it fits you and the more it costs to start fresh.

I believe memory is becoming the main switching cost in AI, and fragmentation comes with it. Whoever holds your context holds you. If four tools each hold a piece, nobody has the whole picture. You don't either.

What can phone number portability teach us?

Telecom had the same problem. In Switzerland, until 2000, if you moved your line to another provider, you got a new number. Then the Federal Office of Communications (OFCOM, BAKOM in German) announced that from March 1, 2000 customers could change providers and keep their mobile, landline and value-added numbers. It said losing your number was an obstacle "many users complained about." It called portability "an important factor for effective competition" (my translation from the German).

The U.S. followed. The FCC required wireless carriers in the 100 largest metropolitan areas to support number portability starting November 24, 2003. Today it says your old provider "cannot refuse to port your number."

To me the point is simple. The thing people use to reach you shouldn't belong to a company you might want to leave. Back then that was your phone number. With AI, it's your context.

Doesn't data protection law already solve this?

Partly. The EU's GDPR gives people a right to data portability. You can receive personal data "which he or she has provided to a controller, in a structured, commonly used and machine-readable format" and pass it to another controller "without hindrance."

I see two gaps. The right covers data you "provided," but much of an AI's memory is inferred from your chats. I'm not a lawyer, but I think it's an open question whether a model's summary of you counts as data you provided. The second gap is practical. A download doesn't get you set up somewhere else. OpenAI lets you request an export of your ChatGPT data. It can take up to 7 days and arrives as a ZIP that "includes your chat history and other relevant account data."

Profile, not memory: what's the difference?

I think "memory" is the wrong word. It describes what a vendor does with your data. It watches, guesses and keeps what it thinks matters. OpenAI's documentation says ChatGPT "decides which available information is relevant" and that the memory summary "may omit details." That's fine for a feature. I don't want my identity to work that way.

A profile is something you write and own. Assistants can suggest additions, and you approve them. It has a clear structure. It lives outside any one product, and many tools can read it.

That last part only became realistic recently. Anthropic introduced the Model Context Protocol (MCP) as "a universal, open standard for connecting AI applications to external systems." According to Anthropic, it has been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot, Visual Studio Code and other products. In December 2025, Anthropic donated it to the Agentic AI Foundation under the Linux Foundation, which it co-founded with Block and OpenAI, to keep it "a neutral, open standard."

So we may get something like number portability without needing a regulator. Your context sits in one place, and the assistants you choose read from it.

What are the seven rules for portable AI context?

I'd measure any AI vendor or profile service against these rules. They started as Eidari's principles. They only mean something if they apply to everyone, us included.

1. You see what an AI gets before it gets it. You can strike out anything.
2. Health, faith and politics stay out unless you add them.
3. You can cut off any AI or delete everything at any time. No forms, no waiting.
4. You see every change an AI makes, and you can undo it.
5. Your data is encrypted. It is never sold, and there are no ads.
6. You can export everything in a plain format.
7. Any assistant can read it through an open protocol, with your permission.

If a product breaks one of these, your context isn't really portable.

What are the honest counterarguments?

Aren't vendors already making switching easier?

Some are. Claude's import tool is a real step, and Anthropic deserves credit for it. But copying once with a prompt doesn't keep anything in sync. By the next day, the two copies start to drift apart.

Isn't a separate profile just one more place my data lives?

Yes, and I think this is the strongest objection. A portable profile means you have to trust whoever hosts it. "Trust us" isn't enough. That's why the seven rules above matter. If a profile service can't meet them, I wouldn't use it.

Isn't the AI's memory better than what I'd write myself?

Often it is. So let assistants write to the profile too. You just need to see every change and be able to undo it.

Don't most people use one AI and never switch?

Maybe. Even then, being able to leave keeps a vendor honest. And if you use several tools, you need one place where your context lives.

Where is this heading?

This part is my opinion. I don't have data for it.

For 2027 I expect something modest. The profile becomes the base layer. Power users will expect a new AI tool to read the context they already have. A tool that only works with what it stores itself will feel like a phone that couldn't take your number.

Owning your profile also changes who you can share it with. Take dating. Today a dating app knows you from a few photos and a short bio. With your permission, it could read the part of your profile you choose to share: your values, what you care about, how you live. That could lead to better matches. You would decide what the app sees, and you could take it back at any time.

Further out, I think it gets bigger. A profile could hold more than your projects. It could hold your values, the way you talk and your story. Then a personality becomes something that can be kept and carried forward. Combined with robotics, I can imagine this helping in care and social work. People could deal with someone familiar instead of a stranger. Grandchildren could get to know a digital version of a grandparent they never had enough time with.

I want to be careful here. This only works if the person clearly agrees, while they are still able to, and decides what is kept and who may use it. It raises hard questions about dignity and grief, and about where a memory ends and an imitation begins. We should answer those before the technology arrives. I'd rather have that conversation about profiles people own than about profiles a platform owns.

What should users and builders do now?

If you use AI tools, ask each assistant now and then to write out what it remembers about you, and keep a copy. Claude's help center suggests asking it to "write out your memories of me verbatim." Then read it. You'll see what it believes about you.

If you build AI products, think of user context as something you borrow. Meet the seven rules. Win on quality, and don't make leaving hard.

Tools will keep changing. I want my context to stay the same across all of them, and I want it to be mine.