ChatGPT's memories are 20 sentences. Your conversation history is thousands. Here's how to bring all of it with you.

If you've been using ChatGPT daily for a couple of years, you've built up something valuable without really thinking about it: context. The AI knows who you are, what you work on, how you like your answers, and what projects you're in the middle of.

That knowledge doesn't just live in the "Memories" feature. Most of it is scattered across hundreds or thousands of conversations.

So what do you do when you want to switch?

Claude has published a guide for migrating from ChatGPT. The short version: paste a prompt into ChatGPT that exports its stored memories, then import them into Claude.

Decent starting point, but the memories are just the tip of the iceberg. ChatGPT's "Memories" are typically 20-50 short sentences. My export had a couple dozen. My actual conversation history, though, contained several thousand conversations over three years. That's where all the real context lives -- your projects, your frustrations, the tools you reach for, how you actually talk, what you've corrected, what you've asked for again and again.

Moving just the memories is like moving to a new house and only bringing the nameplate on the door.

A better method: analyze your entire conversation history

The idea is simple. Instead of exporting only the stored memories, you export everything and then use Claude to pull out what matters.

Step 1: Export from ChatGPT

Go to Settings > Data Settings > Export Data. You'll get a ZIP file with your full history, including chat.html -- a complete HTML file with every conversation you've ever had.

You'll receive a download link via email. It can take several hours depending on how much you've used ChatGPT.

Step 2: Let Claude Cowork analyze the file

If you've used ChatGPT for a few years, your history will be far too large for Claude or any other LLM to handle in a regular chat window. This is where Cowork (or Claude Code) comes in.

Create a new folder on your machine, copy the chat.html file into it, open Cowork, and navigate to that folder.

Select Opus 4.6 as the model, then paste this prompt:

I'm switching LLM services and need to export all context from ChatGPT.

Analyze the file thoroughly and produce one summary file (.md) with synthesized, portable context -- things a new AI assistant needs to know about me.

Don't dump raw conversations, questions, or quotes from individual chats. Focus only on findings and facts worth remembering.

## What to include
### 1. Custom Instructions (verbatim)
All versions of my "Custom Instructions" (both "About Me" and "How to Respond"), reproduced verbatim in code blocks with dates.

### 2. Personal details (synthesized)
Summarize based on all conversations. Don't list individual chats, summarize facts:
- Name, age, location, family situation
- Partner/family (names, profession, details that have come up)
- Job, role, employer/company
- Interests, hobbies, lifestyle
- Health, exercise, habits

### 3. Instructions and preferences
All instructions I've given about tone, style, format, language, and things I've asked to avoid. Preserve my own words where possible. Also include corrections and frustration where they reveal a preference (e.g. "this was too detailed", "wrong again, I'll ask Gemini").

### 4. Projects, goals, and recurring themes
Synthesize from conversations -- don't list individual chats, but describe:
- Companies I run/work with and their context
- Projects and initiatives I've worked on
- Strategies and goals I've discussed
- Recurring themes (what do I use ChatGPT for most?)

### 5. Tools and technology
An overview of tools, platforms, programming languages, and services I use or have used, with approximate frequency/context.

### 6. Location over time
If ChatGPT has injected location data into conversations (system context with "The user's location is..."), list unique locations with dates.

## What NOT to include
- No full conversation index (list of all conversations)
- No activity over time (conversation count per month)
- No raw quotes from conversations unless they express a preference or instruction
- No content from transcriptions/voice recordings (speaker 1/speaker 2 format)
- No questions I've asked -- only what the context reveals about me
- Don't group findings by individual conversations -- group by theme/category

## Format
Each entry should have a date where available, in the format:
- [YYYY-MM-DD] -- content
The file should be concise and scannable. Target: 2-5 pages, not 50.

Step 3: Review the output

Claude will go through all your conversations and produce a structured Markdown file with summarized context. My file ended up around five pages with the stuff that actually mattered: who I am, how I like my answers, what I've been working on, which tools I use, and yes, what my partner's favorite animal is.

Go through the result and adjust. Remove things that are wrong or irrelevant. The goal isn't to import as much as possible into Claude -- it's to import what's most relevant. Add anything that's missing. It's your context. You know best what matters.

Step 4: Import into Claude

From here, follow Claude's official method:

1. Go to Settings

2. Navigate to Capabilities

3. Select Start Import under Memories

4. Paste and select Add to memory

This gets you a much better starting point in Claude than just pulling out old memories with a simple prompt.

It still won't feel the same as ChatGPT from day one, though. You're getting a fraction of what was spread across all those conversations and projects, and it takes a while before Claude catches up. But it goes faster than you'd think. Once you're there, you have memories and context in both places, so switching between ChatGPT and Claude stops being a cold start.

Why this works better

ChatGPT's memory feature stores explicit facts -- things it has "decided" to remember. But a lot of what makes an AI assistant useful is implicit: patterns in what you ask, corrections you give, tools you mention in passing, projects that keep coming up.

Analyzing the full conversation history captures all of that. You can't get ChatGPT (or any other service) to do this kind of analysis with a single prompt. There's too much data for a chat window to process.

Cowork can, though. It reads the file on your machine and boils everything down into a structured document. You end up with a clean context profile you can actually use going forward -- not just in Claude, but anywhere.