Put your ChatGPT archive where your notes already live.

The download is a folder, not a single file, and its shape is fixed — so you can point a script at it later if you ever want to.

Every note opens with the same YAML block. It is there so that Dataview queries and the search bar have something real to work with, not to look thorough:

• title — the conversation title, quoted, so a colon inside it cannot break the block.
• source — chatgpt.
• created and updated — the source timestamps, ISO 8601. Left out entirely when the export carries none, rather than backfilled with the date you happened to run the conversion.
• model — when the export names one.
• messages — the count, including any sitting off the main path.
• conversation_id — the original id, so a note can always be traced back to the export it came from.
• branch_count and messages_off_path — how many alternate branches exist, and how many messages are on them.
• tags: [ai-chat, chatgpt] — so the whole archive is one query away.

This is the part most converters get wrong, and the failure is invisible when they do.

ChatGPT does not store a conversation as a list of messages. It stores a tree. Every time you edited a prompt and regenerated, the abandoned branch stayed in the file, and a single current_node pointer marks the end of whichever path you were last looking at. A converter that walks the JSON in key order will cheerfully interleave three discarded drafts into one conversation and hand it back to you looking completely normal.

Here the current path is resolved first and everything off it is counted. branch_count and messages_off_path in the frontmatter tell you it exists at all. You can include the alternate branches in the note under an Alternate branches heading, or leave them out — and if you leave them out, each omitted message still gets a row in skipped.csv saying so. What does not happen is a silent flatten.

Assets present in the ZIP are written into assets/ under a name derived from their contents, and linked relatively, so the vault survives being moved or renamed.

Assets that are referenced but absent — routine in real exports — are not linked at all. A link to a file that is not there looks identical to a link that works, right up until you click it. The missing asset gets a note in place and a row in skipped.csv. Nothing is fetched from a remote URL to paper over the gap.

Copy conversations/ and assets/ into a vault you already use, or open the whole folder as a new vault on its own. Links and attachment paths are relative, so both work, and nothing needs a plugin to render — it is ordinary Markdown with ordinary YAML.

If you keep the archive as its own vault, index.md is the note to open first.

At the end of a conversion you get three numbers: messages found in the export, messages rendered, and messages skipped. Rendered plus skipped equals found — or the tool says so, in those words, instead of showing you a success screen.

That is the whole point of the thing. A converter that quietly drops what it cannot parse produces an archive that looks complete and is not, and you will not find out for years.