🧠 NFTS Memory MCP

One shared memory for claude.ai, Claude Desktop, and Claude Code — on any computer. Save something in one, it's there in all of them.

Self-hosted, SQLite-backed, zero recurring cost. Runs as one systemd service on your MCP Hub, same pattern as your other custom MCP servers.

How it works

                        ┌─────────────────────────┐
  claude.ai ───────────▶│                          │
  (connector, OAuth)    │   nfts-memory-mcp        │
                        │   ─────────────────      │
  Claude Desktop ───────▶│   Express + MCP SDK      │
  (MCP config, OAuth)    │   SQLite + FTS5 search   │──▶ data/memory.db
                        │   Built-in OAuth server   │
  Claude Code ──────────▶│   Web dashboard          │
  (MCP config, OAuth)    │                          │
                        └─────────────────────────┘

• One process, one database. Every client talks to the same SQLite file over HTTP — there's nothing to sync, because there's only ever one copy.

• 9 MCP tools let Claude save, search, update, and organize memories: memory_save, memory_search, memory_recent, memory_context, memory_update, memory_archive, memory_delete, memory_list_projects, memory_stats.

• Full-text search via SQLite FTS5 (porter stemming, ranked results, snippet highlighting) — fast even at tens of thousands of memories.

• Built-in OAuth 2.0 server (with PKCE + dynamic client registration) so claude.ai, Desktop and Code can each connect as their own authorized client — this is what makes it usable as a claude.ai connector, not just a local tool.

• Web dashboard at your domain root — browse, search, manually add/edit/archive memories, see stats, and manage which apps are connected.

• Projects & categories. Every memory is tagged with a project (nfts, aural, dantectl, general, ...) and a category (fact, decision, preference, todo, config, contact, idea), so Claude can filter context to what's relevant.

Related MCP server: zime-memory

Deploy

On your MCP Hub (or any box with Node 18+):

# 1. Copy this whole folder to the server, then:
cd nfts-memory-mcp
sudo bash deploy/deploy.sh

The script installs a nftsmemory system user, copies files to /opt/nfts-memory-mcp, installs dependencies, and sets up the systemd service. It prints the remaining manual steps (edit .env, nginx + certbot, start the service) — takes about 5 minutes.

Suggested subdomain: memory.nftech-apps.net, following your existing mqtt. / mcp. pattern.

Once it's running:

curl https://memory.nftech-apps.net/healthz
# {"ok":true,"service":"nfts-memory-mcp"}

Verifying the full flow

scripts/verify.sh runs an end-to-end test against a running server (client registration → OAuth login → token → save/search a memory). Point it at your server with BASE_URL if testing remotely, otherwise it defaults to localhost.

Connect each client

All three go through the same OAuth flow — sign in with the ADMIN_USER / ADMIN_PASSWORD you set in .env.

claude.ai (web / mobile)

Settings → Connectors → Add connector → paste https://memory.nftech-apps.net/mcp → Connect → sign in. Done — every conversation on any device now has access.

Claude Desktop

Settings → Connectors → Add custom connector → same URL, same login. (Desktop already shares claude.ai's account memory — this adds Code-style tool access on top, so Desktop can search and save memories explicitly, not just passively recall.)

Claude Code (CLI)

claude mcp add --transport http nfts-memory https://memory.nftech-apps.net/mcp

It'll open a browser for the same OAuth login. After that, every claude session on every machine you run this on shares the same memory.

Check Connections in the web dashboard afterwards — you'll see all three listed with their own token, so you can revoke any one independently (e.g. if you get a new laptop and want to kill the old token).

Using it day to day

You don't need to do anything special — just talk normally:

• "Remember that the Xero email endpoint is POST /Invoices/{id}/Email" → Claude calls memory_save

• "What did we decide about the WiFi upgrade proposal pricing?" → Claude calls memory_search

• "Catch me up on dantectl" → Claude calls memory_context with project="dantectl"

Claude decides when to save/search based on the conversation — you can also be explicit ("save this to memory", "search memory for...") if you want to force it.

The dashboard (https://memory.nftech-apps.net/) is for you: browsing everything, manually adding a memory, archiving stale ones, and seeing which apps are connected.

Maintenance

Backups — it's one SQLite file. Add to your existing weekly Vultr snapshot routine, or simply:

sqlite3 /opt/nfts-memory-mcp/data/memory.db ".backup /path/to/backup-$(date +%F).db"

Logs:

journalctl -u nfts-memory-mcp -f

Updating code: re-run deploy/deploy.sh (it won't touch your .env or data/), then systemctl restart nfts-memory-mcp.

Revoking a lost/compromised device: dashboard → Connections → Revoke next to that client's token.

Why this instead of claude-mem

claude-mem only hooks into Claude Code's session lifecycle — it can't reach claude.ai or Desktop, and it doesn't sync across machines on its own. This is a plain MCP server, so it works identically everywhere Claude can connect to an MCP server, self-hosted on infrastructure you already run, with no per-seat or per-token SaaS cost.

Tech stack

Node.js, Express, @modelcontextprotocol/sdk (Streamable HTTP transport), better-sqlite3 (WAL mode + FTS5), zero build step — it's plain JS, matching the rest of your MCP Hub services.