I don't think AI memory should be a loyalty programme you can't leave. A chatbot remembering your business is useful. Discovering that the easiest way to keep that knowledge useful is to keep paying the same supplier is rather less charming.
By AI memory, I mean the details an assistant carries between conversations: what your business does, how you like things written, which projects matter. You don't have to explain yourself from scratch each time. That's a real convenience, not just another button wearing a sparkly hat.
But I assume the official story about personalisation is also a sales pitch. The company selling the assistant benefits when starting again somewhere else feels exhausting. My read is that memory can become a switching cost without anyone needing to lock the door.
The cancellation button can work perfectly. It's the prospect of explaining your entire business again that keeps you in the chair.
Starting again with a confident stranger who needs the whole business explained sounds less like software migration and more like a networking breakfast. At least the breakfast comes with a croissant.
The moving van isn't the house
Claude has a memory import feature. You can ask another AI service to write out what it remembers, paste that into Claude, and let Claude extract the key information. It also offers a way to export its own memory.
That's useful. It isn't a guarantee that the whole relationship packs neatly into a suitcase. The import feature is experimental, and Claude's own guidance warns that imported memories might not be successfully incorporated.
ChatGPT has a different wrinkle. Its memory summary is a high-level view, not necessarily everything it can draw on. Memory itself doesn't retain every detail from every conversation. Asking what it remembers is sensible; treating the answer as a complete inventory is a leap.
There's an equally revealing distinction in ChatGPT's account-transfer workaround. Exported conversations can be uploaded into a conversation in another personal account as reference material. That doesn't recreate the old chats or transfer memories and settings. This is an account-to-account workaround, not a promise about moving to a rival.
An archive is not a working assistant. A removal van full of kitchen cupboards doesn't mean anyone's fitted the kitchen.
I'm not claiming either company is deliberately sabotaging departures. These are different systems handling different kinds of context. I am saying that an export button and an easy exit aren't the same product feature, however reassuringly similar they look in a sales conversation.
Keep the useful bit outside
For a business owner, I think the answer is deliberately boring: keep a short, readable brief somewhere you control, separate from whichever assistant happens to be fashionable.
Imagine a small furniture retailer. The useful brief might explain who its customers are, what its delivery promise actually covers, which product claims need checking, and what a good customer reply sounds like. Add the current priorities and examples of work someone has approved.
That's an illustration, not a claim about a particular retailer. The point is that those decisions belong to the business. They shouldn't exist only as something a chatbot may or may not have absorbed during a long Tuesday.
I'd rather see a clear page of current instructions than a heroic dump of every conversation. Old arguments, abandoned plans and somebody's inspired suggestion to rename the company aren't automatically useful context. Sometimes they're just evidence that a meeting should have ended earlier.
This isn't an argument for copying confidential material into every service. A portable brief should contain only what you're authorised to share with the destination. Portability doesn't magically turn a customer file into a party invitation.
Then the comparison between assistants becomes more honest. Give each the same brief and the same ordinary task. Judge what comes back, including how much correction it needs. Don't confuse the familiar assistant's head start with proof that its underlying model is better.
Familiarity can be valuable. It just deserves its own line in the reckoning, rather than being smuggled in as intelligence.
Who gets paid for knowing you?
OpenAI benefits if useful accumulated context makes ChatGPT harder to leave. Anthropic benefits if its import tool makes trying Claude less of a chore. Neither needs a secret plot. The commercial incentives are standing in the doorway with a clipboard.
I like the direction of memory imports. I don't think they remove the need for a proper handover. The test is whether the next assistant can do useful work with what you've brought, not whether a box accepted the paste.
For anyone building business software around AI, that's an interesting place to compete: keeping a company's instructions clear, current and usable across suppliers. Not another personality quiz for the chatbot. Something that preserves the decisions the business has actually made.
My judgement is simple. The assistant should earn its place through the work it does now, not the inconvenience of replacing it.
If your AI knows your business better than your own handover notes do, the clever bit may be the retention strategy.
Who owns the relationship when the customer can't take the context with them?