Anthropic / Claude

The mappings below reflect observed export structures as of February 2026, verified against a real export containing 90 conversations and 5134 content parts. Anthropic 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 Claude Settings → Privacy → Export Data
2. Download the ZIP file
3. Extract to find conversations.json, memories.json, projects.json, users.json

Export files

File	Description
conversations.json	Array of conversation objects with chat_messages
memories.json	User memories — conversations_memory (string) and project_memories (dict of UUID to text)
projects.json	Projects — uuid, name, description, prompt_template, docs, creator
users.json	Account info — uuid, full_name, email_address, verified_phone_number

What you get

Full conversation history, user memories (conversations + project-specific), account info, and projects. Claude is one of the most complete exports available.

Field mappings

Conversation-level mapping

Provider field	PAM field	Transform
uuid	provider.conversation_id	direct
name	title	direct
summary	raw_metadata.summary	preserve in raw_metadata
created_at	temporal.created_at	direct (already ISO 8601)
updated_at	temporal.updated_at	direct
account.uuid	provider.account_id	direct
—	provider.name	hardcode "claude"

Message-level mapping

Provider field	PAM field	Transform
chat_messages[].uuid	provider_message_id	direct
chat_messages[].uuid	id	generate UUID or use original
chat_messages[].sender	role	"human"→"user", "assistant"→"assistant"
chat_messages[].text	content.text	direct
chat_messages[].content	content.parts[]	if non-empty, map to multipart (see content type mapping below)
chat_messages[].created_at	created_at	direct (already ISO 8601)
chat_messages[].updated_at	raw_metadata.updated_at	preserve in raw_metadata
chat_messages[].attachments	attachments	map each to Attachment object
chat_messages[].files	attachments	merge with attachments array
—	parent_id	null (Claude conversations are linear)
—	children_ids	[] (Claude conversations are linear)

Content type mapping

Claude’s content[] array contains structured blocks with a type field. Five types have been observed:

Provider content[].type	PAM ContentPart.type	Key fields	Transform
text	text	text, citations[]	Map text directly. Map citations[] to message-level citations[].
thinking	text	thinking, summaries[], cut_off	Map thinking to text. Set is_thought: true on the message. Preserve summaries and cut_off in raw_metadata.
tool_use	—	name, input, id	Map to message-level tool_calls[] with name, input, id.
tool_result	—	tool_use_id, name, content[], is_error	Map to a tool role message. Nested content[] may contain knowledge type with title, url — map to citations[].
token_budget	—	(no useful data)	Discard. Internal Claude token management.

All content parts may include start_timestamp, stop_timestamp (per-part timing) and flags (observed always as null). Preserve in raw_metadata if needed.

tool_use example

{
  "type": "tool_use",
  "name": "web_search",
  "input": { "query": "..." },
  "id": null,
  "message": "Searching the web"
}

Maps to PAM tool_calls[]:

{
  "name": "web_search",
  "input": { "query": "..." },
  "id": null
}

tool_result example

{
  "type": "tool_result",
  "tool_use_id": null,
  "name": "web_search",
  "content": [
    {
      "type": "knowledge",
      "title": "Page Title",
      "url": "https://example.com"
    }
  ],
  "is_error": false
}

Nested content[].type: "knowledge" items map to PAM citations[]:

{
  "title": "Page Title",
  "url": "https://example.com"
}

Memories mapping

memories.json contains a single-element array with the following fields:

Provider field	PAM target	Transform
conversations_memory	memories[] (type: context)	Parse string into individual memories. This is a single block of text — split heuristically or store as one memory.
project_memories	memories[] (type: project)	Each key is a project UUID, value is structured text with sections (Purpose, Current state, Key learnings, Tools). Parse sections or store as a single memory per project.
account_uuid	owner.id or cross-reference	Links to users.json[].uuid

Common import bugs

• The message array field is chat_messages, not messages — this is a frequent source of errors
• Sender values are "human" and "assistant", not "user" and "assistant" — normalize "human" to "user"
• Both attachments[] and files[] exist on messages and may contain different data — merge both into PAM attachments[]

Linear conversations

Claude conversations do not branch. No DAG traversal is needed. Set parent_id: null and children_ids: [] for all messages. The summary field (AI-generated overview) is not present on all conversations — preserve in raw_metadata.

SDK Converters

PAM import support for Claude 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.