Switching AIs? Your Chat History Can Come With You

A better model ships. You read the benchmarks, you try it for an afternoon, and it's genuinely better for what you do. Then you think about the two years of conversations sitting in the app you're already in — the architecture you argued through, the debugging session that finally cracked it, the way it learned to write in your voice — and you stay.

That hesitation is the whole ballgame. It isn't loyalty and it isn't really about the model. It's that starting over means explaining yourself from scratch to something that has never met you.

The switching cost was never the model

Models are the most replaceable part of this stack. They get better every few months, they leapfrog each other, and moving between them takes about five minutes of adjusting to a different tone.

What doesn't move is everything that accumulated around the model. The context. The corrections you made a hundred times until it stopped making that mistake. The decisions you and it reached together, with the reasoning attached. That pile gets more valuable every week, and it lives inside one vendor's app.

Nobody designed this as a trap. It's just what happens when the thing you build up is stored by the same company that sells you the subscription. But the effect is the same: the longer an AI is useful to you, the more expensive it gets to leave.

Built-in memory doesn't come with you

Every major assistant now has some version of memory, and they're genuinely useful. They are also, without exception, local to the product that made them. ChatGPT's memory is a ChatGPT feature. Claude's projects are a Claude feature. There is no transfer button, and there is no reason there would be.

There's a second, quieter problem. Most built-in memory doesn't keep what you said — it keeps a summary of what you said. That's efficient, and it's lossy in a specific way: the exact phrasing of a decision, the constraint you mentioned once, the reason you rejected the obvious approach — those are the first things a summarizer throws out, and they're usually the parts worth keeping.

So even if a transfer button existed, what crossed over would be a sketch of your history rather than your history.

What you can actually take

Here's the part that gets lost in the lock-in conversation: the raw material is already yours, and both major vendors will hand it to you today.

• ChatGPT — Settings → Data controls → Export data. You get an email with a zip; inside it is conversations.json, containing every conversation, verbatim.
• Claude — Settings → Account → Export data. Same idea, same filename.

Do this before you need it. Do it before you cancel anything. An export you already have is worth considerably more than an export you meant to take.

An export is not a memory

Then you open the file, and the limits of owning your data become obvious. It's a few hundred megabytes of nested JSON. You cannot read it. You cannot search it in any way that matters — Ctrl+Ffinds a word, not the conversation where you worked out why that approach wouldn't scale. And no AI can reach into it, because it's a file sitting on your disk.

Data portability gets you a box of your own stuff. It doesn't get you a memory. For the export to be worth anything it has to become searchable by meaning, and it has to be reachable by whatever AI you're using now — including the ones that don't exist yet.

That's the gap Engram fills. It stores conversations complete and verbatim, indexes them for semantic search, and exposes them over MCP — an open protocol that a dozen AI tools already speak. Your history stops being a feature of one app and becomes a layer the apps connect to.

Moving your history in three steps

The whole migration is about ten minutes, most of it waiting.

1. Export. Request the export from whichever assistant holds your history, using the settings path above. The email usually arrives within a few minutes, though a very large account can take longer.

2. Import. Unzip it and drop conversations.json into the import panel on your Engram dashboard. ChatGPT and Claude exports are both accepted and the format is detected for you — there's nothing to configure. If you'd rather stay in the terminal:

engram import ~/Downloads/conversations.json

Add --dry-run to see exactly what would be imported before anything is written. Imports are also fingerprinted against the source conversation, so running the same import twice adds only what's new instead of duplicating your archive — which means you can re-export later and top up.

3. Connect the new AI.Point your new assistant at Engram. It's a one-click connector for ChatGPT and Claude, and a couple of lines of config for Claude Code, Cursor, and the rest.

Then ask the new one something only the old one should know. That moment — a model you started using this morning answering from a conversation you had last spring — is the entire point.

What changes once your history is portable

The obvious benefit is that switching stops hurting. The more interesting one is that you stop having to choose at all.

Most people who work with these tools daily aren't loyal to one anyway. They research somewhere with good web access, write code somewhere with a good agent, and think out loud wherever the writing feels best. Today that means three assistants with three disconnected memories, and a lot of pasting context between them. With one shared memory underneath, it's three interfaces onto the same accumulated context.

And the next time something genuinely better ships, the question is just whether you like it — not what it costs you to start over. Keep the history that took years to build. Change the model whenever you want. Those should never have been the same decision.