AI Diary MCP Server

✨ Highlights

• 🔒 Truly private. Every memory lives in one SQLite file on your disk. Nothing is ever uploaded.

• 🧠 No embeddings, still smart. Retrieval uses SQLite FTS5 + BM25 with a trigram tokenizer, so search is instant and works great for English and 中文 / CJK.

• 🤝 Borrows your platform's brain. When the host supports MCP sampling, recall expands & re-ranks and reflect auto-summarizes — with graceful fallback when it doesn't. The server itself never ships a model.

• 🕸️ A real knowledge graph. Link memories and traverse the web of context.

• 🗂️ Collections, tags, importance, reflection, dedup. Organize like Notion — locally and programmatically.

• 🎨 A beautiful local dashboard. Browse, search, and add memories in your browser (dark & light).

• 🧩 Works everywhere MCP does. Claude Desktop, Cursor, Windsurf, Cline, VS Code, Zed…

Related MCP server: OmniHub

🖥️ The dashboard

A polished local web UI over the same database — npm run dashboard, then open http://localhost:4178.

Browse · full-text search (English + 中文) · filter by tag/collection · spot near-duplicates · add memories · dark/light theme · one-click git snapshot.

🏆 Why it's better than Notion

🚀 Quick start

git clone https://github.com/pangxianggang/ai-diary-mcp.git
cd ai-diary-mcp
npm install
npm run build      # produces dist/index.js — the stdio MCP server
npm run smoke      # self-test (26 assertions)

Try the dashboard with some sample data:

AI_DIARY_DB_PATH=./demo.db node scripts/seed-demo.mjs
AI_DIARY_DB_PATH=./demo.db npm run dashboard      # http://localhost:4178

Override the port with AI_DIARY_PORT.

🎨 The design philosophy

The server is the hippocampus — fast storage, structure, and retrieval. Your platform's model is the cortex — it reasons, summarizes, and decides what to remember.

The server never calls an LLM and never goes online. That keeps it private, free, and trivially portable.

💾 Where memories are stored

A single SQLite file:

Override with AI_DIARY_DB_PATH. Back up by copying the file. Keep it inside a git repo and call snapshot for versioned memory history.

🔌 Connect it to your AI

Use the absolute path to the built dist/index.js.

Claude Desktop

claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "ai-diary": {
      "command": "node",
      "args": ["/absolute/path/to/ai-diary-mcp/dist/index.js"]
    }
  }
}

Cursor / Windsurf / Cline / Zed / VS Code

All use the same shape — command: "node" with args pointing at dist/index.js. Add an env block to relocate the database:

{
  "mcpServers": {
    "ai-diary": {
      "command": "node",
      "args": ["/absolute/path/to/ai-diary-mcp/dist/index.js"],
      "env": { "AI_DIARY_DB_PATH": "/absolute/path/to/my-memory.db" }
    }
  }
}

Restart the client after editing its config.

🧰 Tools

Resource memory://recent (recent entries as Markdown) · Prompt recall-about (recall + summarize a topic).

🧱 Capabilities by phase

1. MVP — SQLite + FTS5 (CJK trigram) + full CRUD over stdio, one-line client config.

2. Smart retrieval — optional MCP-sampling query expansion + re-ranking in recall, automatic fallback to plain FTS5.

3. Structure — tags, collections, links / knowledge graph traversal, importance, soft-forget.

4. Reflection — reflect auto=true consolidates memories via the host model; find_duplicates for de-duplication.

5. Experience — a polished local web dashboard + Markdown export + optional git-versioned history.

🔍 How search works (without embeddings)

1. FTS5 + BM25 ranks entries by relevance. The trigram tokenizer indexes 3-character sequences, which is what makes substring and CJK search work.

2. Metadata filters (tags, category, collection, time range, importance) narrow results.

3. Queries shorter than 3 characters fall back to a LIKE substring scan.

4. With smart=true, a sampling-capable host expands the query and re-ranks candidates — but the core works fully without it.

🛠️ Development

npm run dev        # tsc --watch
npm run typecheck  # type-check only
npm run smoke      # in-memory end-to-end check (26 assertions)
npm run dashboard  # local web UI

📄 License

MIT