Memory Vault — Persistent Context for Claude (Any Account)

A self-hosted memory layer for Claude. Your context lives as plain markdown files (Obsidian-compatible) instead of inside any one Claude account, so switching accounts no longer means starting over.

Tested and working: vault read/write/search, and the full MCP protocol round-trip (initialize, tools/list, tools/call) over both transports below.

How it works

vault/*.md  (plain markdown + YAML frontmatter — open directly in Obsidian)
     ↑↓
src/vault.js          (save / get / search / list / delete)
     ↑↓
src/mcp-tools.js       (wraps vault.js as 5 MCP tools)
     ↑↓
src/stdio-server.js    src/http-server.js
(local, Claude Desktop)  (hosted, any Claude account, any device)

Five tools are exposed to Claude:

• save_memory — store a fact/decision under a topic

• get_memory — retrieve everything under one topic

• search_memory — fuzzy search across all topics

• list_memory_topics — overview of what's stored

• delete_memory_topic — remove a topic (destructive)

Related MCP server: SAE4U Memory

Setup

npm install
cp .env.example .env

You picked "both" earlier, so here's both modes:

Mode A — Local (Claude Desktop only, simplest)

Runs on your machine, talks to Claude Desktop over stdio. No network, no deployment, no auth needed — Claude Desktop launches the process itself.

1. Find your Claude Desktop config file:

   • Windows: %APPDATA%\Claude\claude_desktop_config.json

   • Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

2. Add this entry (use the absolute path to this project):

{
  "mcpServers": {
    "memory-vault": {
      "command": "node",
      "args": ["C:\\path\\to\\mcp-memory-vault\\src\\stdio-server.js"],
      "env": {
        "VAULT_DIR": "C:\\path\\to\\mcp-memory-vault\\vault"
      }
    }
  }
}

3. Restart Claude Desktop. The tools above will be available in any conversation. This config is per-install of Claude Desktop, not per-account — so if you're logged into different accounts in the same Desktop app, this stays connected across all of them already.

Mode B — Hosted (every account, every device, including claude.ai web)

This is the one that actually solves "switch accounts, keep context," since it's not tied to any single Desktop install.

Deploy to Render (free tier):

1. Push this folder to a GitHub repo.

2. On Render: New → Web Service → connect the repo.

3. Build command: npm install

4. Start command: npm run start:server

5. Add an environment variable: MCP_SHARED_SECRET = some long random string (generate one with node -e "console.log(require('crypto').randomUUID())").

6. Deploy. Render gives you a URL like https://your-app.onrender.com.

Connect each Claude account to it:

1. In claude.ai (or Claude Desktop) → Settings → Connectors → Add custom connector.

2. URL: https://your-app.onrender.com/mcp

3. Add a custom header: x-mcp-secret = the same MCP_SHARED_SECRET value.

4. Save. Repeat once per account (one-time per account, as discussed — after that it's permanent).

Free-tier note: Render's free web services spin down when idle and take ~30-50s to wake on the first request after inactivity. If that's annoying, the keep-warm trick you already know from FLUX (a cron ping every 10-14 min via cron-job.org or UptimeRobot, hitting /health) fixes it.

Using the Obsidian connection

VAULT_DIR can point anywhere — including an actual Obsidian vault folder. If you do that, every memory Claude saves shows up as a normal markdown note in Obsidian immediately, and anything you edit in Obsidian is immediately what Claude reads next time. No sync step, because it's literally the same files.

VAULT_DIR=/path/to/your/Obsidian/Vault/claude-memory

Testing it works

With the server running (npm run start:server), from another terminal:

curl http://localhost:8080/health

Should return {"ok":true,"vaultDir":"..."}.

What this does NOT do (be aware)

• Doesn't sync to ChatGPT/Gemini/Grok automatically — those need a different bridge (Custom GPT Action / browser extension), discussed separately.

• search_memory is fuzzy keyword search (via Fuse.js), not semantic vector search. Good enough for a personal vault at this scale; if it ever feels too literal, swapping in embeddings is a contained change to vault.js only.

• Single-writer assumption: fine for one person across many accounts, not built for concurrent multi-user writes.