memex-md-mcp

You like Obsidian? Your LLM will love it too.

Memex: Vannevar Bush's 1945 concept of a "memory extender" - a device for storing and retrieving personal knowledge. The conceptual ancestor of personal wikis and second brains.

MCP server for searching and navigating markdown vaults. Point it at your Obsidian vault (or any markdown folder) and get semantic search, wikilink/backlink traversal, and note renaming with automatic link updates.

What memex is: A search and navigation layer over your markdown files. SQLite with FTS5 for keyword search, embeddinggemma for semantic similarity, wikilink graph for backlinks.

What memex isn't: An automatic memory system. It won't capture context or write notes for you. For that, check out claude-mem (automatic memory compression with hooks and summaries). Memex pairs well with workflow layers on top—see my agent workflows for an example using memex as the knowledge backend.

Quick Start

claude mcp add memex -- uvx --from 'memex-md-mcp==1.*' memex-md-mcp

Then ask Claude to help configure your vaults - it has mcp_info() which explains everything. Or manually edit your settings (see Configuration below).

Version note: The above pins to the latest 1.x release for stability. For bleeding edge, use memex-md-mcp@latest—but watch the repo for releases, since major bumps may require deleting your index (~/.local/share/memex-md-mcp/memex.db).

What This Does

Memex gives Claude read access to your markdown vaults. It creates a local index at ~/.local/share/memex-md-mcp/memex.db and logs to ~/.local/share/memex-md-mcp/memex.log. The index contains:

• Full-text search index (FTS5) for keyword matching
• Embeddings (google/embeddinggemma-300m) for semantic similarity
• Wikilink graph for backlink queries
• Extracted frontmatter (aliases, tags)

On each query, memex checks file mtimes and re-indexes any changed files.

Note: Initial indexing requires embedding computation. Example: ~3800 notes took ~7 minutes on an RTX 3070 Ti. Subsequent queries only re-index changed files and are fast.

Hidden directories (.obsidian, .trash, .git, etc.) are excluded from indexing.

Writing to notes happens through Claude Code's normal file tools.

Configuration

Add to ~/.claude/mcp.json (global) or .mcp.json (per-project):

{
  "mcpServers": {
    "memex": {
      "command": "uvx",
      "args": ["memex-md-mcp@latest"],
      "env": {
        "MEMEX_VAULTS": "/home/user/knowledge:/home/user/project/docs"
      }
    }
  }
}

Multiple vault paths are colon-separated. Project .mcp.json overrides global config entirely (no merging), so list all vaults you need.

Optional: Disable Semantic Search

If you only need wikilink navigation and keyword search (no GPU/embeddings):

"env": {
  "MEMEX_VAULTS": "...",
  "MEMEX_DISABLE_SEMANTIC": "1"
}

When disabled: search() only works with keywords, explore() returns empty similar list.

Tools

search(query?, keywords?, vault?, limit=5, page=1, concise=True) — semantic search over vaults.

• query: Describe what you're looking for in natural language. Use 1-3 sentences, question format works well. If omitted, runs FTS-only mode with keywords.
• keywords: Optional list of exact terms to boost. Required if query is omitted.
• page: Page number for pagination (1-indexed).
• concise: Returns only paths by default. Use concise=False for full content.

search("What authentication approach did we decide on? I remember we discussed OAuth.")
search("How does the caching layer handle invalidation?", keywords=["Redis", "TTL"])
search(keywords=["PostgreSQL"])  # FTS-only mode

explore(note_path, vault, concise=False) — graph traversal from a note.

Returns outlinks (what it references), backlinks (what references it), and semantically similar notes not yet linked. Includes full content of the explored note (not neighbors). Outlinks include image embeds (![[image.png]])—use Read tool to view them.

note_path can be a full path or just the title (if unique in vault):

explore("api-design", "/home/user/vault")              # by title (if unique)
explore("architecture/api-design", "/home/user/vault") # by path

Typical workflow: search() to find entry points → explore() promising results to read content + see connections.

rename(note_path, new_name, vault) — rename a note and update all wikilinks.

Renames the file and updates all [[wikilinks]] pointing to it. Handles edge cases:

• Path-based links: [[subdir/note]] → [[subdir/newname]]
• Title-based links: [[note]] → [[newname]]
• Preserves aliases/headings: [[note#section|Display]] → [[newname#section|Display]]
• Ambiguous links (multiple files share a name): skipped with warning

rename("old-name", "new-name", "/home/user/vault")
rename("docs/guide", "manual", "/home/user/vault")  # also updates [[docs/guide]] links

mcp_info() — returns this README.

Workflow Integration

Add to your project's CLAUDE.md (adapt paths to your setup):

# Memex MCP

You have access to markdown vaults via memex. Use them to find past work, discover connections, and document knowledge that helps future sessions.

Vaults:
- ...

Search tips:
- Use 1-3 sentence questions, not keywords: "How does the auth flow handle token refresh?" beats "auth token refresh"
- Mention key terms explicitly in your query
- For exact term lookup, use keywords parameter with a focused query
- For precise "find this exact file/string" needs, use grep/rg instead — memex is for exploration

Workflow: search() returns paths by default (concise) → explore() promising results to read content + see connections → Build context before implementation.

For how I use memex, see my agent stuff.

Benchmarks

Performance:

• For now mostly my own vibes, still developing a proper workflow around this.
• So far I only tested semantic and FTS search in isolation on my 3.8k note Obsidian vault to tune it.

Speed:

• Initial indexing: ~7 minutes for ~3800 notes (RTX 3070 Ti)
• Subsequent queries: ~instant

Development

uv sync
make check          # ruff + ty
make test           # pytest
make release-patch  # 0.2.6 -> 0.2.7, tag, push
make release-minor  # 0.2.6 -> 0.3.0
make release-major  # 0.2.6 -> 1.0.0