How do I open conversations.json from a ChatGPT export?
conversations.json is one large JSON array where each conversation stores its messages as a tree under "mapping", so a text editor shows it as one unreadable line.
Updated November 15, 2026
What is in the export
- conversations.json: every conversation, in one array. Each item has a
title, acreate_time, anupdate_time, acurrent_nodeand amapping. - chat.html: a page that shows the same conversations in your browser.
- Image and attachment files with names that start with
file-, sometimes inside subfolders. user.json,message_feedback.jsonand a few other small files. You can ignore them for this job.
What chat.html gives you, and its limits
Double-click chat.html and your browser shows every conversation, one after another, on a single page. For a short history that is enough to read and to search with Ctrl+F or Cmd+F.
- It is one page for everything. With a long history it loads slowly or not at all.
- There is no file per conversation, so you cannot move one chat into your notes or give it to another tool.
- It is not Markdown. Copying text out of it loses the structure of code blocks and headings.
Read conversations.json by hand
- Unzip the export.
- Do not open
conversations.jsonin a normal text editor. Use a tool made for large JSON. On the command line,jq '.[].title' conversations.jsonlists the titles andjq '.[0]' conversations.jsonprints the first conversation with line breaks. - In one conversation, read
current_node. That is the ID of the last message shown in ChatGPT. - Find that ID in
mapping. Each entry has aparent, a list ofchildrenand amessage. - Follow
parentfrom entry to entry until there is no parent. Reverse that list. This is the conversation in reading order. - For each
message,author.rolesays who wrote it (user,assistant,systemortool) andcontent.partsholds the text.
Things that trip people up
- The messages are a tree, not a list. Every time you edited a prompt or regenerated an answer, the old version stayed in
mappingas another branch. Reading the entries in file order mixes the branches. - Times are Unix seconds.
1794700800means November 15, 2026, 00:00 UTC. - Not every message is text.
content.content_typecan betext,code,execution_output,multimodal_textand others. Code is incontent.text, not inparts. - Images are pointers. A part such as
file-service://file-AbC123refers to a file in the ZIP whose name starts withfile-AbC123. Some of those files are not in the export. - Some messages are hidden. System messages and entries marked
is_visually_hidden_from_conversationnever appeared in the chat. - The file can be too big for a script that loads it whole. A script that reads the file into one string fails when the file is larger than the language's string limit or the free memory.
The faster way
ConvoTidy does these steps for the whole export in your browser and writes one Markdown file per conversation, with the images and an index. Nothing is uploaded. It is free for your 25 most recent conversations and $9 one-time for all of them.