How to Export Claude Conversations, and What the Seven Files Hold
By Matthew Firth, founder of ModelBrain · Oct 30, 2026 · 5 min readOur own proposal: treat a Claude export as a checklist of parts and not as a single file, because the conversations and the memory arrive in separate downloads. Our real export came as seven zip parts; the conversations part held 272 conversations and 20,089 messages (452 MB unpacked from an 88 MB zip), and the memory sat in its own part with 154 memory files.
Request the export
On the Claude web app or Claude Desktop, open Settings, then Privacy, then Export data. Anthropic emails a download link to your account address. The link expires 24 hours after delivery and you must be signed in to use it. The export cannot be started from the iOS or Android app. On a Team or Enterprise plan, members can export their own data only after an owner turns on "Allow members to export their own data", and a Primary Owner exports the whole organization from Organization settings, then Data and privacy. Anthropic also says an export cannot be imported into another personal Claude account. We read these steps on Anthropic's Export your Claude data page on October 30, 2026.
It is seven downloads, not one
Our own proposal: treat the export as a checklist of parts. When we asked for ours, it came as seven separate zip files listed in one manifest, and the memory was in a different part from the conversations.
| Part | What it held in our export |
|---|---|
| conversations | One conversations.json, 452 MB unpacked from an 88 MB zip |
| memories | One JSON with account_uuid, conversations_memory and memory_files: 154 memory files, each with a path, content and an updated time |
| projects | Eight project JSON files, 9.2 MB |
| frames | 856 entries under artifacts/<id>/, each with artifact.json, page.md and comments.json, 138 MB |
| design_chats | Ten JSON files |
| claude_code_cloud | routines.json and routine_runs.json |
| light_metadata | users.json and login_history.json |
Keep login_history.json out of anything you share. It holds your sign-in history.
What is inside conversations.json
The file held 272 conversations and 20,089 messages: 9,636 from the human and 10,453 from the assistant. A conversation has the keys account, chat_messages, created_at, name, summary, updated_at and uuid. A message has attachments, content, created_at, files, parent_message_uuid, sender, text, updated_at and uuid. The content of a message is a list of typed blocks.
| Block type | Count in our file | Share of 214,136 |
|---|---|---|
| tool_use | 74,371 | 35% |
| tool_result | 74,336 | 35% |
| thinking | 38,334 | 18% |
| text | 26,032 | 12% |
| injected_prompt_block | 390 | under 1% |
| image | 362 | under 1% |
| document | 311 | under 1% |
Only the text blocks are what you and Claude said to each other. About 70 percent of the file is tool calls and their results, which is why an export of 20,089 messages unpacks to 452 MB. A further 3,180 messages carry no content blocks at all, and 49 of the 272 conversations contain no dialogue a reader could use.
The four conversations that are bigger than the rest
Four of our conversations were 37.5, 38.0, 54.6 and 64.8 MB each, because tool results and attachments sit inside them. Any tool that refuses a single record over 32 MiB will skip those four without telling you. If you move an export into a memory or search tool, check how it handles one very large conversation before you trust its count.
Download each part once
Our first export's links had all expired by the time we came back to them, so we requested a second. For the second one we signed in, took each part once, one at a time, and kept the original zips untouched. Count the parts against the manifest before you close the email.
Where to go next
The same exercise for the other two assistants is in ChatGPT, Claude and Grok export formats compared, and the ChatGPT file is a tree, not a transcript, as that article shows.