What actually transfers
Here is what is documented, in the vendors' own words, and it is mostly about time and deletion.
Deletion is all-or-nothing, and fast. X's help centre is specific: "You can delete all of your conversation history at once. Deleted conversations are removed from our systems within 30 days, unless we have to keep them for security or legal reasons." The control lives in settings, where you will see "Delete Conversation History" and then a confirmation to "Delete your interactions, inputs, and results."
There is no documented per-conversation pruning in that flow. It is one action covering everything.
The retention clock is the same on the way out. SpaceXAI's privacy policy states it twice, for two different paths. On deletion: "if you choose to delete any or all of your conversations or if you choose to delete your account, we will delete the data within 30 days unless it is necessary to retain the data for legal, compliance, or safety purposes." And on private sessions: "when Private Chat is turned on, conversations will not appear in your conversation history and your conversations will be deleted from SpaceXAI systems within 30 days unless it is necessary that they be kept longer for legal, compliance, or safety purposes."
So anything you had in Private Chat was already on a timer, whether or not you were planning to leave. SpaceXAI also notes that "the length of time we retain data may depend on the features or settings you use," which is a reminder that the answer differs by how you have been working.
Your own copy is a request, not a button. SpaceXAI's privacy policy describes the rights route: "You may have the right to request to access, correct, update or delete your personal information, subject to certain applicable legal exceptions." Separately, X documents an archive download — Settings, then Download your data, then Request data, after which you get an email and can download a .zip. X's help page describes what that archive is for in one line: "Downloading your X archive allows you to browse a snapshot of your X information, starting with your first post." That page documents the request-and-download steps; it does not describe Grok conversation contents. So treat the archive as something to request and then actually open and inspect, rather than as a known-good backup of your chats.
Feedback changes the picture in one specific way. X notes that "you can voluntarily submit feedback on a conversation even if you opt out of model training," and that if you do, "X and xAI may use that conversation and any such conversation related data to train and fine-tune Grok and improve your experience." Worth knowing if you have been liberal with the thumbs-up button on conversations you considered private.
What carries to ChatGPT is whatever you wrote down. On the destination side, OpenAI documents where personalization lives — "Memory controls are available in Settings > Personalization > Memory" — and what that memory is: a continually updated synthesis of context from past chats, with legacy saved memories described as "details you've directly told ChatGPT to remember." Neither is an import target for another product's transcripts. You will be re-establishing context by telling it, not by uploading it.
There is one more asymmetry worth naming, because it shapes what is even worth moving. SpaceXAI's own position has been that memory belongs to the individual Bot rather than to the account — the reasoning we examined in Grok's per-Bot memory model. If you built several Bots, what each one knew was scoped to it by design. There is no single account-level knowledge base to hand over.
The manual migration
Step 1: Harvest before you touch any deletion control
Do this first and do not skip to the settings screen. The deletion control is all-or-nothing, the clock is 30 days, and there is no documented undo.
Work through your conversation history and pull out four categories. They are the four that people always regret losing:
Decisions with reasons. Anything where you and Grok worked through a choice and landed somewhere. Write the conclusion and the reason in one sentence each. "We are not using the second API because its rate limit resets hourly, not per-minute" is worth more than the twenty exchanges that produced it.
Standing preferences. How you want things written, formatted, addressed. These accumulate invisibly across months and are the single most annoying thing to rebuild, which is the frustration described in when Grok forgets your personal preferences.
Research threads you are still in the middle of. Not the answers — the state. What you have already ruled out, which sources you checked, what question you were about to ask next. Grok forgetting research context covers why this is the category that degrades fastest.
Anything a Bot knew that you want to keep. Per-Bot scoping means this material never lived anywhere central, so it has to be collected per Bot.
Request the X archive in parallel so it is ready while you work, then open it and see what is actually in it before you rely on it for anything.
Step 2: Rebuild in ChatGPT through the two surfaces it documents, then verify
ChatGPT gives you two honest places to put this, and OpenAI describes the split: "For explicit information or instructions, you can add it to your Custom Instructions. For information shared via conversations, ChatGPT can remember relevant details for you."
Put the standing preferences in Custom Instructions. They are explicit, they apply always, and that is what the surface is for.
For the decisions and research state, tell ChatGPT directly in conversation and let memory pick them up — then verify, because the visible list is not the full picture. OpenAI is candid about this: the memory summary "will not include everything that ChatGPT remembers based on your chats," and "if you want to know if ChatGPT has remembered something, just ask in chat."
Two more things worth setting up while you are in there. If you want a topic contained, use a Project: OpenAI documents that with project-only memory, "chats can reference other conversations in the same project, but they cannot reference conversations outside the project, and chats outside the project cannot reference conversations in it." And know the opt-out: "Temporary Chats do not use existing memories or create new memories."
The catch, and it is the same catch every time: you are now maintaining this by hand in a second product. Getting the standing facts in is a good hour's work. Keeping them accurate as they change is the part nobody plans for, which is why the general problem of assistants feeling forgetful survives the move.
The Better Way: Write the conclusions somewhere a 30-day clock cannot reach
The lesson of this particular migration is not about Grok. It is that when your project's reasoning lives inside a chat history, your project's reasoning inherits that history's retention policy, deletion controls and product roadmap. A switch that clears everything in one action is a perfectly reasonable privacy feature. It is a terrible place to keep the reason you chose one architecture over another.
So keep the transcripts wherever they are and stop treating them as the record. Hold the conclusions in a layer that no single assistant owns, and the next move — to ChatGPT, or away from it later — becomes a change of interface rather than a rescue operation.
That layer is what MemoryLake holds: the decisions, preferences and constraints you would otherwise be re-typing into whichever assistant you are using this month.
Step 1: Create an API key
Sign in, open your workspace settings, and generate an API key. This is the credential your assistants use to read the same layer, so create it once and keep it available wherever you work.
Step 2: Upload your first memories
Start with the four categories from the harvest: decisions with reasons, standing preferences, open research state, and anything a Bot knew that matters. One fact per entry, short enough to act on without opening a document.
Step 3: Connect your AI & agents
Connect ChatGPT and anything else you use. The same preferences and decisions arrive in each, so a fresh session starts informed instead of starting over.
What this changes in practice
The first difference is that deletion becomes safe. Clearing a conversation history is a privacy action you should feel free to take. When the durable part is recorded elsewhere, it stops being a knowledge-destroying action as well.
The second is that Private Chat becomes usable. A mode whose conversations are deleted within 30 days is genuinely useful for sensitive work, and it is unusable as a workspace if it is also where your conclusions live. Separate the two and you can use the private mode for what it is good at.
The third is that the per-Bot boundary stops being a wall. Scoping memory to a Bot is a defensible design and we said so when SpaceXAI argued for it. But architectural decisions, terminology and ownership are facts every Bot needs. Holding those in one shared layer means each Bot can stay narrow without each Bot being ignorant.
The fourth is that you stop paying the re-establishment tax at every switch. The steady drift of context across sessions described in when ChatGPT loses context between sessions is the same tax, charged by a different vendor.
Best practices for leaving a chat-history-based assistant
Harvest before you delete, always. All-or-nothing controls plus a 30-day removal window means there is no partial recovery. Assume the history is gone the moment you confirm.
Open the archive you requested. A .zip you never unzipped is not a backup. X's page describes the archive as a snapshot of your X information starting with your first post; find out what that means for your account rather than assuming.
Write conclusions, not transcripts. Forty exchanges compress to two sentences. Those two sentences are the asset; the forty exchanges are evidence you will almost never re-read.
Treat feedback buttons as a disclosure. X documents that submitting feedback on a conversation may allow that conversation to be used for training even if you opted out of model training generally. That is a reasonable trade for improving results; it is worth making knowingly.
Verify the destination took it. OpenAI recommends asking in chat whether something was remembered, because the summary does not show everything. Do that for your five most important facts rather than trusting the list.
Keep sensitive material out of both. X's guidance on Grok is direct — "Please do not share personal data or any sensitive and confidential information in your conversations with Grok" — and the same caution applies to anything you put in a memory store. Record the decision, not the credential. Our note on backing up your ChatGPT memory covers the same discipline in the other direction.
Conclusion
Grok's documented controls are good at what they are designed for: letting you clear your history and knowing it is gone within 30 days. They are not designed to hand you a portable copy of what you learned, and the archive route that exists is a request whose contents you should inspect rather than assume.
Which leaves the migration in your hands, in the most literal sense. The work is an hour of reading your own conversations and writing down the few dozen sentences that actually mattered. Do that once, into a layer that outlives any single assistant, and it is the last time you will do it — because the next migration will be a connection change instead of an archaeology project.