@remember-md/mcp

Local MCP server for the Remember.md second brain. Run via npx, point any MCP client at it, query your markdown brain semantically.

Status: v0.0.1 — skeleton. Not yet functional. Active development.

What it does

Exposes your local markdown brain (a folder of .md files organised PARA-style by the Remember.md plugin) as a set of MCP tools any MCP client can call — Claude Code, OpenClaw, Cursor, Codex CLI, Claude.ai web, ChatGPT custom GPTs, anything that speaks the Model Context Protocol.

Planned tools:

• search_brain(query, top_k) — semantic + BM25 hybrid + wikilink-expand

• get_file(path) — read a brain file

• list_recent(period, kind?) — recent journal / notes / decisions

• query_persona() — current Persona.md content

• dashboard_snapshot() — counts + top beliefs + active projects

• propose_belief(claim, evidence) — write candidate to Inbox/

How it works

• Storage: node:sqlite (Node 22.5+ stdlib) + sqlite-vec extension for vector search + FTS5 for BM25 — no server, no native compilation, no toolchain.

• Embeddings: @huggingface/transformers running quantized Xenova/bge-micro-v2 (384d, ~17 MB) locally — no cloud calls.

• Sync: on-demand mtime + content-hash incremental reindex at query time. The brain (markdown) is the source of truth; the index in .remember/index.db is rebuildable.

• Graceful degradation: if vector loads fail, falls back to FTS5-only; if both fail, falls back to ripgrep.

Install

You don't install it. Point your MCP client at it via npx:

Claude Code (via the Remember.md plugin's /remember:init)

The Remember.md plugin automatically configures Claude Code's MCP layer to launch this server. Just run /remember:init.

Cursor / Codex / other MCP clients

Add to your MCP config:

{
  "mcpServers": {
    "remember": {
      "command": "npx",
      "args": ["-y", "@remember-md/mcp"],
      "env": {
        "REMEMBER_BRAIN_PATH": "/absolute/path/to/your/brain"
      }
    }
  }
}

First run downloads the package (~15–30s) and the embedding model (~17 MB, one-time). After that, queries are sub-second.

Configuration

Privacy

Local-only. No cloud calls. No telemetry. The brain folder + index never leave your machine. Embedding model runs in-process via ONNX Runtime.

License

MIT — see LICENSE.

Related

• Remember.md plugin — the capture / curate / persona side that produces the brain this server queries

• Remember.md spec — the markdown standard