Index your Grok export in ModelBrain

• Request your export from Grok. On the web, look under Settings, then Data Controls, then Export Account Data. Grok emails or shows a download when it is ready. The wording of Grok's menus may change.
• Use ModelBrain 0.6.6 or later. You can use the zip as it is or unzip it first.
• Nothing is added until you add it. ModelBrain never guesses a source kind from a file extension.

1. Open the ModelBrain menu, choose Vaults, then Add Source, then Grok export….
2. A prompt asks whether your export is a zip file. Choose Yes to pick the zip. Choose No to pick the unzipped folder.
3. ModelBrain indexes it in the background.

From a terminal, connector_sync register <id> <path> --type grok-export does the same. For example: connector_sync register grok ~/Downloads/grok-export.zip --type grok-export. The path can also be the folder that holds the zip, or the prod-grok-backend.json file itself. Add --kind-option include_do_not_remember only if you want the temporary chats described below.

• Only the one chat-history file, prod-grok-backend.json. The zip is never loaded whole, so the images do not cost memory.
• One memory per conversation, keyed by its conversation id. The id does not depend on which folder holds the conversation, so a fresh export of the same account leaves unchanged conversations alone.
• The text of your messages and Grok's replies. A reply with no text that made images becomes the placeholder [image]. No image data is read.
• Each turn sits under a User or Assistant heading, so search can tell who said what. A passage you wrote ranks above the same words in a reply.
• If two exports overlap, the copy with the later modified time wins.

• Side branches. Only the chain from the last reply back to the start is indexed.
• Steps, thinking traces, tool cards, web search results and image links.
• Replies that failed with an error and no text.
• Temporary chats. They are skipped and counted. Registering with --kind-option include_do_not_remember includes them.
• Every other file in the export, such as account and billing files and the image assets. They are never opened.

A container file over 4 GiB is refused and recorded as a failure, not dropped. Set MODELBRAIN_CONTAINER_MAX_BYTES to change that ceiling. A single conversation over 32 MiB is refused the same way. The chat-history file is held in memory while it is read, so one over 256 MiB uncompressed is refused.

ModelBrain removes a conversation from memory only after it has read the export to the end. If the file cannot be parsed, every conversation already indexed stays put until the next good pass.

1. Ask your assistant to run ModelBrain diagnostics. The source reports its kind, members, members_failed and last_container_bytes, and the number of temporary chats left out.
2. Ask a local assistant about something you once discussed with Grok.