Switching AI Models? Take Your Memory With You

September 2026 has been a switching month. OpenAI shipped GPT-6 Astra, then Sol and Luna. Anthropic shipped Claude Opus 5.5. Every leaderboard reshuffled, and a lot of people are trying a different assistant for the first time in a year.

Switching the model takes ten seconds. Switching your context is the hard part: the chats you learned from, the preferences the old assistant remembered, the documents you kept pasting in. Leave those behind and the shiny new model starts out knowing less about you than the one you left.

Here’s what each vendor actually lets you move today, and the setup that makes the question irrelevant.

What the vendors let you move (September 2026)

Gemini has the most complete import. Since March, Google’s switching tools can pull your memory summary and full chat history from ChatGPT and Claude, from ZIP exports up to 5GB. Limits: no project files, attachments or generated images, only prompts and responses. And the tool is not available in the EU/EEA, the UK or Switzerland.

Claude can import and export its memory. Anthropic’s help center documents both directions, so you can carry what Claude knows about you to another tool, or seed Claude with memories from elsewhere.

ChatGPT exports everything (Settings → Data controls → Export data), but at the time of writing there is no memory import. Bringing Claude or Gemini memories into ChatGPT means telling it, one fact at a time.

Chat history exports from all three arrive as a ZIP by email, sometimes after days, in JSON or HTML that is readable by machines more than by people.

So you can move some of it, in some directions, in some countries. That’s not portability. That’s a series of one-way doors.

The three things you actually lose

When people say “I don’t want to lose my ChatGPT”, they usually mean three different things:

1. The memory summary. The short list of facts about you. Useful, but small, and the easiest to rebuild.
2. The conversations. The long threads where you worked something out: a strategy, a refactor, a research deep-dive. This is the real value, and the hardest to move.
3. Your sources. The articles, docs and threads you pasted in. They were never really “in” the assistant; they were in your clipboard for five minutes.

Vendor import tools mostly handle #1, partly handle #2, and don’t touch #3.

The setup that makes you model-agnostic

The fix is to stop treating any assistant as the place where your knowledge lives. Keep it in plain Markdown files you own, and let whatever model you’re using this month read them.

Markdown is the right format for this because every model reads it natively, it’s tiny compared to HTML or PDF, and it’s just text. It opens in any editor, syncs in any cloud folder, and will still open in twenty years.

In practice:

1. Save the conversations that matter as Markdown. When a chat produces something worth keeping, save it as a .md file: every turn, with code and artifacts as proper code blocks. Minibase does it in one click on claude.ai, chatgpt.com and Gemini: open the conversation, click the extension, and you get a clean file. Guides: Claude and ChatGPT.

2. Save your sources the same way. The article you’d have pasted in, the docs page, the Reddit thread: one click turns each into a Markdown file with its URL and date.

3. Keep them in one folder per topic. Minibase Vault organizes saves into knowledge bases, which are plain subfolders on your disk. “Pricing research”, “Rust migration”, “Trip to Japan”.

4. Write your own memory file. One about-me.md with the facts you want every assistant to know: your role, your stack, your writing style, what to avoid. You control every line, and it’s safer than letting an assistant infer them (see AI memory poisoning).

5. Connect the folder to whichever model you use. Minibase Vault connects to Claude through MCP with one toggle, so Claude reads your files directly. For any other assistant, drag the relevant files in. They’re just Markdown.

What switching looks like after that

Next time a new model tops the charts, you don’t migrate anything. You open the new assistant, give it your about-me.md and the folder for the project you’re on, and it starts where the old one left off, with your actual sources instead of a summary of them.

That’s the real lesson of a month with two flagship launches: the models are going to keep changing. The layer worth investing in is the one that doesn’t, your own knowledge, in a format every model can read.

Related reading

• GPT-6 Astra vs Claude Opus 5.5: filling a 1M context
• How to save a Claude conversation as Markdown
• Markdown: the language of AI in 2026