library

Library

A read-only window into the user's Claude Desktop conversation history. Each chat lives on Anthropic's servers. The library MCP server pulls them through Claude's session cookies and caches them locally for fast outline and full-text search.

State lives in ~/.claude/library/library.db. Use the library_* tools for every operation. Never read or write the cache file directly.

When this skill activates

Two paths.

1. Awareness path. The user references their own Claude history in natural language. Examples:

   • "what did I tell Claude about that bug?"
   • "find the chat where I designed the schema"
   • "pull up my conversation about pricing"
   • "didn't I work through this with Claude already?"

   Take the find sequence below.

2. Explicit path. The user typed /library or /library <action>. Jump to Actions.

Don't fire for things obviously about the current Claude Code session. Don't fire for Claude Code session history (that's a different MCP).

Find sequence

Default loop when the user wants to recall something from past chats.

1. Search first. Call library_search with a tight query. 2-4 keywords, FTS5 syntax. Don't dump the whole user query in.
2. Read the snippets. Pick the most relevant 1-3 hits. Hits come in sections: [human]/[assistant] from messages, [doc] from project knowledge-base files, [memory #N] from memory snapshots, [artifact] from code blocks.
3. Drill in. For any hit with a UUID, call library_open <uuid> to read the full item. For conversations, call library_outline first, then library_open with from/to bounds.
4. Pull only what's needed. Don't fetch the full transcript when 5 messages will do.
5. Answer the user. Quote the relevant passage. Cite the conversation title or doc filename, plus date.

If library_search returns nothing, call library_list to see what's cached. The user may need to sync.

Language-aware querying

FTS5 doesn't stem, translate, or fuzzy-match. The cached corpus may be in a different language than the question, or mixed.

Check library_list titles first to spot the corpus language.

• Titles consistently in one language: search in that language.
• Mixed: search both variants with OR and merge.

If the question and the titles are in different languages, translate the salient nouns into the corpus language before searching. People ask in their working language. The conversation was written in whatever they used at the time.

Hit rate goes up when you:

• Strip stopwords. "the broken build" becomes broken.
• Drop quotes on multi-word phrases. Quotes force exact adjacency and miss often.
• Pick the distinctive noun. Names, jargon, and proper nouns beat verbs and adjectives.
• OR the noun across two languages when in doubt.

When a title looks right but search misses, fall back to library_list for browsing or library_outline on the suspected UUID.

Actions

(no argument) or status

Call library_status. Print the result. If cached conversations: 0, suggest /library sync.

sync [--full] [--org NAME]

Call library_sync. Set full: true if --full is present. Pass org if given. Print the result.

Sync is incremental by default. It pulls conversation metadata for every page and only fetches message bodies for new or updated conversations. --full re-fetches everything.

list [N] [--project NAME]

Call library_list with limit and optional project. Default limit 25. Print the table.

search <query> [--kind KIND]

Call library_search with the query. Optionally pass kind to restrict to one source:

• all (default): messages, docs, memory, artifacts
• messages: conversation message bodies only
• docs: project knowledge-base documents only
• memory: organization memory snapshots (no query = list all snapshots)
• artifacts: extracted code blocks and antArtifact tags
• tool_calls: tool invocations (query matches tool name)
• citations: web citations (query matches title, URL, site name)
• shares: shared conversation snapshots

projects [--name NAME] [--detail]

Call library_projects. With no args, lists all projects with conversation counts. With name=X, filters to one project. With name=X detail=true, returns the full record including system prompt.

outline <uuid>

Call library_outline. Print the outline.

open <uuid> [from] [to]

Call library_open. Detects what kind of item the UUID belongs to (conversation, doc, artifact, share, or memory snapshot) and returns the appropriate content. For conversations, from/to bound the message range.

Unknown action

Print:

Common moves:
  library sync                   pull new and updated conversations
  library list                   show cached conversations
  library search <query>         full-text search across messages, docs, memory, artifacts
  library search <query> kind=X  restrict to one source (messages/docs/memory/artifacts/tool_calls/citations/shares)
  library projects               list projects with conversation counts
  library projects name=X detail=true  full project record including system prompt
  library outline <uuid>         turn-by-turn outline of one conversation
  library open <uuid>            read any item by uuid (conversation, doc, artifact, share, memory)
  library status                 cache and auth check

Don'ts

• Don't library_open an entire 200-message conversation when the user asked one question. Use library_outline plus a bounded library_open.
• Don't ask the user to sync without first checking library_status to see if the cache is empty.
• Don't fabricate conversation titles or UUIDs. If a search returns nothing, say so.
• Don't read ~/.claude/library/library.db with Bash. Use the tools.
• Don't fire on questions about the current session's history.