Switch AI models.
Keep everything.
ChatGPT and Claude both remember now — but neither one's memory moves to the other. Switch providers and you leave it behind. Here's how to switch without starting over.
This is the practical page: what you actually lose when you switch, why it happens, and the fix. For the longer argument about why your continuity layer should be vendor-independent in the first place, see why Memrith is BYOK.
What you actually lose when you switch
Start with what's true today, because it's changed. ChatGPT can reference your past conversations — since April 2025 it reaches across your whole history, not just a handful of saved facts, and a lighter version landed for free users in mid-2025. Claude does its own version: it can search past chats, carry context forward, and keep separate memory per Project, with memory reaching free users in early 2026. The old line — "your AI forgets you every session" — is no longer accurate for either one. They remember.
So the thing you lose when you switch isn't memory in general. It's the specific memory that one provider built about you — and that memory is locked to them. The months ChatGPT spent learning how you write, what you're working on, and which answers you kept don't come with you to Claude. They can't: that context lives in your OpenAI account, in OpenAI's format, on OpenAI's servers. Claude starts without it and rebuilds its own, slowly, from your new conversations. Switch back and the same thing happens in reverse.
Concretely: you've spent six months getting ChatGPT to understand a long-running project — the constraints, the decisions you already ruled out, the way you like things explained. A better model ships on Claude and you want to move. None of that understanding transfers. You either stay put to keep the memory, or you switch and spend the next stretch re-teaching the new model everything the old one already knew. That's the tax: not that the new AI can't remember, but that it has to start over learning you.
Why provider memory doesn't transfer
It's worth being precise about why, because it isn't a missing feature someone will patch. Each provider's memory is part of that provider's product. It lives in your account with them, in their internal format, in their cloud, under their retention and training terms. ChatGPT's memory was never designed to be readable by anyone but ChatGPT. Claude's memory — including the per-Project memory — was never designed to be handed to OpenAI. There's no shared format and no export-to-rival path, in either direction.
That's by design, not by accident. Memory is what makes leaving expensive. The longer a provider remembers you, the more switching costs you — so the incentive runs toward keeping that memory inside the platform, not toward making it portable. None of this requires bad faith from anyone. It's just what happens when the memory belongs to the provider instead of to you.
The heuristic: if the memory that understands your work lives in a provider's account, every switch resets it — no matter how good each provider's memory gets on its own. The fix isn't a better provider memory. It's keeping the memory somewhere a switch can't touch.
The fix: a memory layer you own
Memrith keeps your memory in a layer that sits on your side of the line — a documented file on your own machine — and lets whichever AI provider you've connected read from it. Because the memory isn't inside ChatGPT or Claude, switching between them doesn't move it. There's nothing to migrate. The provider is the part that changes; the memory is the part that stays.
This works because Memrith is BYOK — you bring your own AI
key. You connect your own Anthropic (Claude), OpenAI (GPT), or OpenRouter
key and pay that provider directly for usage, with no Memrith markup.
Switching providers is a Settings change: Settings → AI Provider →
paste the new key, save. Your entries, your accumulated context, and your
edits are unchanged. The next conversation opens with the same memory,
routed through the new model.
The model can change. Your memory doesn't.
Switching the provider swaps which box reads from your memory. The memory itself never moves — it lives on your machine.
How a switch works with Memrith, step by step
Say you've been routing through ChatGPT and you want to move to Claude. The whole switch is a settings change — under a minute, and your memory is never in flight:
- Get a key for the provider you're switching to. Create an Anthropic (Claude) account and generate an API key. If you've only ever used the consumer ChatGPT or Claude apps, this is the one genuinely new step — there's a walkthrough for each provider.
- Open Settings → AI Provider in Memrith. Pick the provider you're moving to and paste the new key. Nothing about your memory is touched here — you're changing the engine, not the file.
- Save. That's the switch. Memrith now routes your conversations through the new model.
- Open a conversation. The new model arrives already knowing your projects, your past decisions, and how you like things explained — because it's reading the same memory the old one was. Nothing to re-teach.
- Switch back any time. Want GPT for one task and Claude for another, or want to go back next month? Repeat the settings change. The memory is constant across every switch.
Switching is not a migration. With provider-held memory, "switching" means rebuilding context on the new platform. With Memrith there's nothing to rebuild — the memory was already yours and already local, so the new provider just starts reading it. The difference is the whole point of this page.
Switching, side by side
(ChatGPT memory → Claude, or back)
(BYOK, memory you own)
The honest trade-off
Two things are real costs, and pretending otherwise would be the kind of copy this page is trying not to be.
First, you still set up BYOK. If you've only used the consumer ChatGPT or Claude apps, generating an API key is a new step — a few minutes per provider, with a walkthrough for each. It's not nothing. It buys you the thing the rest of this page is about: memory that doesn't reset when you change models.
Second, you still pay each provider for usage. BYOK means the AI bill goes to OpenAI or Anthropic directly, on what you actually use, with no Memrith markup. Memrith itself is a one-time $9.99 purchase — not a subscription — but the provider usage is yours to pay. For most people that lands at a few dollars a month, and you can set a hard spend cap in your provider's dashboard.
If "I just want it to work with zero setup, on one provider, and I don't plan to switch" describes you, the first-party memory in ChatGPT or Claude is genuinely simpler at the start, and free. The reason to use Memrith is the day you do want to switch — and your context comes with you instead of staying behind. That's the trade the whole page is built around, and it's laid out at more length in why Memrith is BYOK.
Common questions
Can I move my ChatGPT memory to Claude?
Not directly. ChatGPT's built-in memory lives in your OpenAI account, in OpenAI's format and cloud. There's no supported way to hand that memory to Claude — Anthropic can't read OpenAI's memory store, and OpenAI doesn't export it as something another provider can load. The practical fix is to stop relying on the provider's memory for the context you care about, and keep that context in a layer you own. With Memrith, the memory is a local file on your machine that whichever provider you've connected reads from — so moving from GPT to Claude doesn't move your memory, because your memory was never inside GPT to begin with.
How do I switch AI models without losing my history?
If your history only lives inside one provider's memory, switching providers leaves it behind — there's no transfer between, say, ChatGPT and Claude. To switch without losing it, the context has to sit somewhere both providers can reach. In Memrith, switching is a Settings change: AI Provider, paste the new key, save. Your entries, your accumulated context, and your edits stay exactly as they were, and the next conversation opens with the same memory, routed through the new model.
Does Claude's memory transfer to GPT?
No. Claude's memory — including its per-Project memory — stays within Anthropic, under Anthropic's terms, and doesn't transfer to OpenAI any more than ChatGPT's memory transfers to Claude. The lock works in both directions. It isn't an oversight; provider-held memory is built to keep you on that provider. Memrith sidesteps the problem by keeping the memory on your side of the line, so a switch in either direction changes the model, not the memory.
Do I have to re-explain everything when I change AI providers?
With provider-held memory, effectively yes — the new provider starts without the context the old one had built about you, so you rebuild it through use. With Memrith, no: the memory you've accumulated is yours and unchanged across the switch, so the new model arrives already knowing your projects, your decisions, and your phrasings. You set up the new key once; you don't re-explain who you are.
Can I use one memory across ChatGPT and Claude?
Not with the providers' own memory — ChatGPT's memory and Claude's memory are separate stores in separate accounts that can't share. A memory layer you own can. Memrith keeps one local memory that any connected provider draws on, so the same context is available whether you're routing through GPT, Claude, or OpenRouter today. You bring your own key for each, and the memory stays in one place: yours.