xAI / Grok

The mappings below reflect observed export structures as of February 2026, verified against real export data. xAI does not natively support PAM. These mappings are best-effort compatibility guidance. Provider export formats may change without notice. Importers MUST be versioned.

How to export

1. Go to grok.com → Account settings → Download your data
2. Download and extract the ZIP file

Export files

The ZIP contains files under ttl/30d/export_data/<user_uuid>/:

File	Description
prod-grok-backend.json	Main file: conversations, projects, tasks, media posts
prod-mc-auth-mgmt-api.json	User profile and sessions
prod-mc-billing.json	Credits balance
prod-mc-asset-server/	Uploaded files referenced by asset UUID (code, images, PDFs, etc.)

What you get

Full conversation history with DAG structure, projects with custom personalities, and media posts. Asset files (uploads) are stored as UUID-named directories containing a content file with no extension.

Field mappings

Conversation-level mapping

Conversations are wrapped: each item is {"conversation": {...}, "responses": [...]}.

Provider field	PAM field	Transform
conversation.id	provider.conversation_id	direct
conversation.title	title	direct
conversation.create_time	temporal.created_at	direct (ISO 8601)
conversation.modify_time	temporal.updated_at	direct (ISO 8601)
conversation.user_id	provider.account_id	direct
conversation.starred	raw_metadata.starred	preserve
conversation.system_prompt_name	raw_metadata.system_prompt_name	preserve
—	provider.name	hardcode "grok"

Message-level mapping

Each response is wrapped: {"response": {...}, "share_link": {...}}.

Provider field	PAM field	Transform
response._id	provider_message_id	direct
response._id	id	generate UUID or use original
response.parent_response_id	parent_id	map provider ID → PAM ID
—	children_ids	reconstruct by inverting parent_response_id
response.sender	role	see Role Normalization
response.message	content.text	direct
response.create_time	created_at	BSON → datetime.fromtimestamp(int(v["$date"]["$numberLong"])/1000, tz=UTC).isoformat()
response.model	model	direct
response.cited_web_search_results	citations	map each {url, title, preview} to Citation {url, title, snippet} (preview → snippet)
response.generated_image_urls	attachments	map each URL to Attachment with type: "image"
response.file_attachments	attachments	map each asset UUID
response.web_search_results	raw_metadata.web_search_results	preserve
response.thinking_trace	raw_metadata.thinking_trace	preserve
response.thinking_start_time	raw_metadata.thinking_start_time	preserve
response.thinking_end_time	raw_metadata.thinking_end_time	preserve
response.agent_thinking_traces	raw_metadata.agent_thinking_traces	preserve
response.steps	raw_metadata.steps	preserve (tool execution chain)
response.query	raw_metadata.query	preserve
response.query_type	raw_metadata.query_type	preserve
response.xpost_ids	raw_metadata.xpost_ids	preserve
response.webpage_urls	raw_metadata.webpage_urls	preserve
response.card_attachments_json	raw_metadata.card_attachments_json	preserve
response.error	raw_metadata.error	preserve
response.metadata	raw_metadata.grok_metadata	preserve (renamed to avoid collision)

Role normalization

Four distinct sender values have been observed: "human", "assistant", "ASSISTANT" (uppercase), and model names such as "grok-3". Normalization MUST be case-insensitive. Treat any value that is not "human" (case-insensitive) as "assistant".

Provider value	PAM normalized value
human	user
assistant	assistant
ASSISTANT	assistant
grok-3 (model name)	assistant
any non-human value	assistant

Timestamp handling

Grok uses two different timestamp formats within the same file:

Level	Format	Transform
Conversation	ISO 8601	direct
Message	BSON {"$date":{"$numberLong":"<ms>"}}	datetime.fromtimestamp(int(v["$date"]["$numberLong"])/1000, tz=UTC).isoformat()

Two separate parsers are needed for the same file.

Critical notes

• Wrapper nesting: Every conversation is {conversation, responses} and every response is {response, share_link}. Must unwrap twice.
• DAG structure: parent_response_id is present on most messages. Supports branching conversations. children_ids must be reconstructed by inverting parent references.
• Empty messages: Some responses have message: "" (empty string). These are typically image generation responses where the content is in generated_image_urls.
• Steps = tool use chain: steps[] contains structured tool execution with tagged_text and tool_usage_results. This is Grok’s equivalent of function calling. Preserve in raw_metadata.
• Thinking traces: Two mechanisms — thinking_trace (inline string) and agent_thinking_traces[] (array). Both may be present on the same response. Preserve in raw_metadata.
• Asset files: Uploads are stored as UUID-named directories in prod-mc-asset-server/, each containing a content file with no extension. Use content inspection to determine type.

SDK Converters

PAM import support for Grok is provided by official SDK Converters maintained by the PAM project. See the Importing Guide for general import instructions and the Provider Overview for the full compatibility matrix.