Switching assistants used to mean starting the relationship again. You explained your job, your projects and your tone rules in one app, then typed the whole story again in the next one. Memory changed that only inside each app. ChatGPT remembers you for ChatGPT. Claude remembers you for Claude. Porting memory between ChatGPT, Claude, Gemini and Grok is still a manual job, but two of the four now offer a real import, and a clean file of your own makes the other two far less painful.
This tutorial follows the help pages from OpenAI, Anthropic and Google and the Grok controls documented by xAI, checked on November 11, 2026. It claims no test run of its own. The Verification section tells you how to judge the result in your own accounts.
What You Will Build or Learn
You will end with one portable memory file and one assistant that has loaded it. The file holds the durable facts an assistant should carry: your role, your active projects, the names and terms you use often, and the rules for how you want answers shaped. It does not hold your chat history. No memory import in this tutorial moves your old chats as chats, though Gemini offers a separate chat import noted in Step 5.
The gain is control. The master file is yours to read, edit and reuse. When you switch again, or when memory drifts, you start from that file and not from memory about memory. The same file habit sits behind our tutorial on portable Gemini Skills, where a file you own travels between tools that read a shared standard. Memory has no shared standard yet. A clean text file is the closest thing to one.
The master file is the product of this tutorial. Each import is just a copy of it. Illustration: Dirty Static
Before You Begin
Decide what deserves to travel before you extract anything. Good candidates are facts that stay true for months and shape many answers. Poor candidates are moods, guesses an assistant made about you, and facts tied to one chat and no other.
One limit shapes the whole method. What travels is text that you can read. A summary that one assistant wrote about you is not a full copy of what it can draw on. OpenAI states that the ChatGPT memory summary is a high level view and does not necessarily include every detail or source that Memory can reference. Expect loss, plan to check, and keep the master file as your record of what should have moved.
Step 1: Build the Master File
Action: create a plain text file named memory-port.txt in a folder you control, and paste in the headings below. Leave the sections empty for now. You will fill them from your source assistant in Step 2 and clean them in Step 3.
ROLE AND CONTEXT
Facts about my work and my situation that stay true for months.
ACTIVE PROJECTS
Each project on its own line, with its name and one line on its current state.
PEOPLE AND TERMS
Names, spellings and terms I use often, with one short note on each.
HOW TO ANSWER ME
Tone, length, format and language rules I want in most answers.
FACTS TO REMEMBER
Durable preferences and facts, one per line.
NEVER ASSUME
Claims assistants have made about me that are wrong or out of date.
Expected result: you hold one empty frame with six sections. The last section looks odd and earns its place. Telling a new assistant what not to assume stops the most common import error, where one wrong summary line gets copied forward as truth for years.
Step 2: Pull Memory Out of Your Current Assistant
Action: use only the route for the assistant you use now, and paste the result into the matching sections of your master file. Where a route tells you to run a prompt, run the prompt in a new chat in the source assistant. The import flows in Steps 4 and 5 supply their own prompt, and that prompt also works as an extraction prompt here.
ChatGPT. Open Settings, then Personalization, then Memory. Read the memory summary or the saved memories list that your account shows, and copy the items you want into the file. OpenAI notes that controls vary by plan, region and platform, and that the summary may omit details. Also ask ChatGPT in a chat what it remembers about you, and compare that answer with the settings list. Then copy your custom instructions from the Personalization area into the HOW TO ANSWER ME section, since custom instructions are guidance you wrote and memory is context ChatGPT built. The data export under Settings, then Data controls, then Export data is a separate archive. OpenAI describes the downloaded ZIP as your chat history and other relevant account data, and its export page does not present it as a memory list. Use it as a backup, not as the source for this file.
Claude. Open Settings, then Your Claude, and select Manage next to Your memory to read what Claude holds, or open Settings, then Memory where that layout appears. Copy the entries you want. Anthropic also suggests asking Claude to write out its memories of you verbatim, exactly as they appear in its memory. Save that answer as a source note and compare it with the panel before you trust either one.
Gemini. Copy your Instructions for Gemini, which you can reach through Settings and help, then Personal Intelligence, then Instructions for Gemini, or at gemini.google.com/saved-info in the layouts that use that address. Those instructions are the part you wrote in your own words. Gemini also learns from past chats when Memory is on, and that layer has no single list to copy. Run the extraction prompt in a Gemini chat and treat the answer as a draft to check, in the same way as the ChatGPT answer above.
Grok. Grok memory rests on the Personalize Grok with your conversation history switch in Settings, then Data Controls, a control still marked Beta in the documented settings. Grok has no memory panel and no memory import or export box in the xAI documentation gathered for this tutorial. Run the extraction prompt in a new Grok chat and treat the answer as unchecked draft text. The account data download at accounts.x.ai/data is a full account archive in JSON form. It can serve as a backup of your chats, but it is not a readable memory list and it will not fill your master file on its own.
Expected result: your master file now holds rough text under most headings. It will contain repeats, stale facts and at least one claim you do not recall teaching anyone. That mess is normal. Step 3 exists because extraction is the easy part.
One file, four doors: two doors have an import desk inside, and two need you to hand the folder over in a chat. Illustration: Dirty Static
Step 3: Clean the Master File
Action: edit the file line by line. Keep one fact per line. Delete any fact you cannot confirm from your own knowledge. Move wrong or stale claims into NEVER ASSUME rather than leaving them in a fact section. Shorten project notes to a name and a current state. Cut any line about a person who is not you unless that person would be fine with the line living in four company systems.
Expected result: the file reads as a short briefing that a smart stranger could use on day one. If a line would embarrass you when read back in a sales call, it does not belong in a file built for pasting.
Step 4: Load the File Into Claude
Action: use this step when Claude is the destination. Open Claude on the web or in Claude Desktop and go to Settings, then Your Claude, and select Start import. Where your layout shows the older path, Memory sits under Settings, then Memory, or under Settings, then Capabilities. Copy the prompt Claude provides, run it in your source assistant if you have not yet extracted your memory, then return to Claude, paste your cleaned master file text into the import box in place of the raw answer, and select Add to memory. Anthropic lists the import for Free, Pro, Max, Ultra, Team and Enterprise plans.
Expected result: Claude extracts key points from the pasted text and stores them as single memory entries that you can review in the memory panel. Anthropic calls memory imports experimental and still in active development, and warns that Claude may not always incorporate imported memories with success. It also notes that Claude memory aims at work topics, so personal details with no link to work may not be kept. Read every entry it creates before you call the move done.
Step 5: Load the File Into Gemini
Action: use this step when Gemini is the destination. Go to gemini.google.com, open Settings at the bottom, and select Import memory to Gemini. Copy the prompt Google provides. Run that prompt in your source assistant, copy its response, then return to Gemini, paste the response into the text field, and select Add memory. Google states that after the import, Gemini creates a new chat thread and integrates the content so it can remember the details in later chats.
If you cleaned a master file in Step 3, paste a copy of that cleaned text when the flow asks for the response, and keep the original file untouched outside Gemini.
Expected result: the new chat thread shows the imported content in Gemini. Check it against your file and correct any line that changed shape in transit. Google limits this feature to a personal Google Account and to users who are 18 or over, and lists the European Economic Area, Switzerland and the United Kingdom as places where it is not available. It is also absent from Gemini in Google Messages, Gemini in Chrome and Gemini on Android XR.
Google offers a second, separate move on the same screen. Under Import chats, select Add and upload a ZIP of exported chats from the other platform, up to 5 GB per file and up to five ZIP files per day. Imported chats land in your activity and can be continued in Gemini. That is a chat history move, not a memory move, and this tutorial does not need it for the memory file to work.
Step 6: Load the File Into ChatGPT
Action: use this step when ChatGPT is the destination. OpenAI documents no memory import box for ChatGPT, so work in two parts. First, open Settings, then Personalization, then Custom Instructions where your account shows that control, and paste the HOW TO ANSWER ME section, trimmed to fit. Those instructions guide every answer and are the strongest rules you own in ChatGPT. Second, open a new chat, paste the rest of your cleaned file, and ask ChatGPT to remember the facts in it, one section at a time if the file is long.
Expected result: ChatGPT confirms what it saved, and the saved items begin to appear under Settings, then Personalization, then Memory. Open that page after the chat and read the list. Any fact that failed to land should be stated again in a short plain request, such as asking ChatGPT to remember one named fact about one named project. A pasted file that got a polite reply but no saved entries has not been imported, whatever the chat claimed.
Step 7: Load the File Into Grok
Action: use this step when Grok is the destination. Open Settings, then Data Controls, and confirm that Personalize Grok with your conversation history is on, since Grok personalization rests on your conversation history. Then open a new chat, paste your cleaned file in parts, and after each part ask Grok to remember those facts and to use them in later chats. Keep each paste short enough that Grok can repeat the facts back to you in its reply.
Expected result: Grok repeats the facts back without new errors. Treat that echo as a receipt for that chat, not as proof of a stored memory panel, because xAI documents no panel where you can audit the result line by line. The Verification checks below matter more for Grok than for any other destination, and a miss there means you paste that section again in a fresh chat rather than arguing with the old one.
Verification
Start a new chat in the destination and ask three questions that only the file can answer: one about your role, one about an active project by name, and one that tests a tone rule from HOW TO ANSWER ME. Then ask the assistant what it remembers about you, and compare the answer with the master file, section by section.
In Claude and ChatGPT, also open the memory panel or list described in their steps and read the actual entries. In Gemini, open the new chat thread that the import created and the instructions page, and check both. A cheerful reply in one chat can draw on the pasted text in that chat alone. A correct entry in the settings list is the evidence that memory, and not just context, moved.
Fix every gap in this order. Correct the master file first. Then correct the destination, by editing its entry where a panel allows it or by stating the fact again in a fresh chat. If you fix only the destination, the next port will carry the old error forward again.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| The source answer leaves out facts you know it used | The memory summary is a partial view, as OpenAI states for ChatGPT, or the learned layer has no list to copy | Copy the settings list and the chat answer, merge them by hand, and keep only facts you can confirm |
| No Start import appears in Claude | The account layout differs, or the plan or app version does not show the feature | Check Settings, then Your Claude, then Settings, then Memory and Settings, then Capabilities. Try the web app or Claude Desktop, where Anthropic lists the import |
| Claude kept work facts and dropped personal facts | Claude memory aims at work topics, in its own warning | State the personal fact again in a chat tied to your work context, or accept the loss and keep that fact in your master file for other tools |
| No import option appears in Gemini | The account is a work, school or supervised account, or the region is excluded | Use a personal Google Account if you have one. If you live in an excluded region, load the file by pasting it into a Gemini chat and saving key rules as Instructions for Gemini |
| ChatGPT replied with thanks but saved nothing | A paste in a chat is context for that chat, not a saved memory, until ChatGPT stores it | Ask for one fact at a time with a plain remember request, then read Settings, then Personalization, then Memory to confirm each entry |
| The new assistant repeats an old error with confidence | The source summary carried a wrong claim and the destination copied it as fact | Move that claim into NEVER ASSUME in the master file, delete the wrong entry in the destination, and state the correct fact again |
| Grok seems to forget the file in a new chat | Personalization from conversation history is off, or the pasted chat fell out of the history it draws on | Turn on the personalization switch in Data Controls and paste the core sections again in the new chat, starting with role and projects |
| Imported chats in Gemini will not upload | The file is over the size limit or is not a ZIP | Split the export as the source tool allows, keep each ZIP at 5 GB or under, and stay within five ZIP files per day |
Next Steps
Keep the master file where you keep other project records, and update it when a project ends or a rule changes. A two minute edit to that one file saves a fresh extraction next time, and extractions only get less complete as summaries stack on summaries.
When a skill or a set of standing instructions proves itself in one assistant, store it next to the memory file in the same folder. Together they form a small porting kit: who you are, what you are working on, and how you want the work done, in text that no single company controls. That kit, updated a few times a year, is what makes the next switch a chore of minutes rather than a week of small corrections.
Facts checked against OpenAI Help Center pages on memory in ChatGPT and on exporting ChatGPT history and data, the Google Gemini Apps Help page on importing from other AI platforms, and reporting of the Anthropic Help Center article on importing and exporting Claude memory, on November 11, 2026. Grok routes were checked against xAI settings documented for grok.com and the xAI account data page. This procedure was verified against those documents and was not presented as a personal test run.
Sources
- OpenAI, Help Center: Memory in ChatGPT
- OpenAI, Help Center: Exporting your ChatGPT history and data
- Anthropic, Help Center: Import and export your memory from Claude
- Google, Gemini Apps Help: Import from other AI platforms to Gemini Apps
- xAI, Grok FAQ and grok.com Data Controls: Personalize Grok with your conversation history, and Export Account Data
- Illustrations: Dirty Static
END OF TRANSMISSION
Stay curious. Question everything.
