add-skill-claude-mem

Add Claude-Mem to Agent Containers

Run all commands automatically. Only pause if a step fails.

UX Note: When asking the user questions, prefer using the AskUserQuestion tool instead of just outputting text.

How It Works

The claude-mem worker daemon runs on the host (port 37777, bound to 0.0.0.0) and stores observations in a SQLite + vector DB at /root/.claude-mem/. Docker containers reach the host via host.docker.internal using the --add-host=host.docker.internal:host-gateway flag (configured in container-runtime.ts).

Preflight

Before installing, verify NanoTars is set up:

[ -d node_modules ] && echo "DEPS: ok" || echo "DEPS: missing"
docker image inspect nanoclaw-agent:latest &>/dev/null && echo "IMAGE: ok" || echo "IMAGE: not built"
if grep -q "ANTHROPIC_API_KEY\|CLAUDE_CODE_OAUTH_TOKEN" .env 2>/dev/null || [ -f "$HOME/.claude/.credentials.json" ]; then echo "AUTH: ok"; else echo "AUTH: missing"; fi

If any check fails, tell the user to run /nanotars-setup first and stop.

Prerequisites

• claude-mem plugin installed (claude plugin add @thedotmack/claude-mem)
• Bun runtime installed

Install

1. Verify prerequisites:

   ls /root/.claude-mem/claude-mem.db
   ls /root/.claude/plugins/cache/thedotmack/claude-mem/*/scripts/worker-service.cjs 2>/dev/null | head -1
   BUN_PATH=$(which bun 2>/dev/null || echo "/root/.bun/bin/bun")
   $BUN_PATH --version
   

   If the database does not exist, stop and tell the user to install claude-mem first.

2. Kill orphaned worker processes:

   pkill -f 'worker-service.cjs.*--daemon' 2>/dev/null || true
   sleep 2
   curl -s --max-time 2 http://127.0.0.1:37777/api/health >/dev/null 2>&1 && echo "WARNING: Port 37777 still in use" || echo "Port 37777 is free"
   

3. Create wrapper scripts at /root/.claude-mem/:

   /root/.claude-mem/run-worker.sh:

   #!/bin/bash
   PLUGIN_DIR="/root/.claude/plugins/cache/thedotmack/claude-mem"
   WORKER=$(ls -td "$PLUGIN_DIR"/*/scripts/worker-service.cjs 2>/dev/null | head -1)
   if [ -z "$WORKER" ]; then
     echo "Error: claude-mem worker script not found in $PLUGIN_DIR" >&2
     exit 1
   fi
   echo "Starting worker from: $WORKER"
   exec /root/.bun/bin/bun "$WORKER" start
   

   /root/.claude-mem/stop-worker.sh:

   #!/bin/bash
   PLUGIN_DIR="/root/.claude/plugins/cache/thedotmack/claude-mem"
   WORKER=$(ls -td "$PLUGIN_DIR"/*/scripts/worker-service.cjs 2>/dev/null | head -1)
   if [ -z "$WORKER" ]; then
     curl -s -X POST http://127.0.0.1:37777/api/admin/shutdown 2>/dev/null
     exit 0
   fi
   exec /root/.bun/bin/bun "$WORKER" stop
   

   Make executable: chmod +x /root/.claude-mem/run-worker.sh /root/.claude-mem/stop-worker.sh

4. Create systemd service for claude-mem worker at /etc/systemd/system/claude-mem-worker.service:

   [Unit]
   Description=Claude-Mem Worker Daemon
   After=network.target
   
   [Service]
   Type=oneshot
   RemainAfterExit=yes
   ExecStart=/root/.claude-mem/run-worker.sh
   ExecStop=/root/.claude-mem/stop-worker.sh
   WorkingDirectory=/root/.claude-mem
   Environment=HOME=/root
   Environment=PATH=/root/.bun/bin:/root/.local/bin:/usr/local/bin:/usr/bin:/bin
   
   [Install]
   WantedBy=multi-user.target
   

5. Enable and start service:

   systemctl daemon-reload
   systemctl enable claude-mem-worker
   systemctl start claude-mem-worker
   sleep 3
   curl -s http://127.0.0.1:37777/api/health
   

6. Test connectivity from a container:

   docker run --rm --add-host=host.docker.internal:host-gateway \
     --entrypoint node nanoclaw-agent:latest \
     -e "fetch('http://host.docker.internal:37777/api/health').then(r=>r.json()).then(d=>console.log('OK:',JSON.stringify(d))).catch(e=>console.error('FAIL:',e.message))"
   

7. Add environment variable:

   grep -q "^CLAUDE_MEM_URL=" .env 2>/dev/null || echo "CLAUDE_MEM_URL=http://host.docker.internal:37777" >> .env
   

8. Copy plugin files:

   cp -r ${CLAUDE_PLUGIN_ROOT}/files/ plugins/claude-mem/
   

9. Plugin Configuration: By default this plugin is available to all groups and channel types. To restrict access, edit plugins/claude-mem/plugin.json and set:

   • "groups" to specific group folder names (e.g., ["main"]) instead of ["*"]
   • "channels" to specific channel types (e.g., ["whatsapp"]) instead of ["*"]

   Ask the user if they want to restrict access. Most users will keep the defaults.

10. Rebuild and restart:

    npm run build
    nanotars restart
    

Verify

Tell the user:

Setup is complete. Test it by sending a WhatsApp message like "remember that my favorite coffee is a flat white" and then in a new conversation ask "what's my favorite coffee?"

Version Management

The wrapper scripts dynamically resolve the plugin version at runtime. After upgrading:

systemctl stop claude-mem-worker
pkill -f 'worker-service.cjs' 2>/dev/null; sleep 2
systemctl start claude-mem-worker

Troubleshooting

• Worker not starting: journalctl -u claude-mem-worker -f and check /root/.claude-mem/logs/
• Container can't reach host: Verify --add-host=host.docker.internal:host-gateway is in container-runtime.ts extraRunArgs()
• Port conflict: lsof -i :37777 -- kill orphaned workers first
• Service status: systemctl status claude-mem-worker

Per-group credential overrides: Not applicable. Claude-mem is a system-wide service shared across all groups.

Uninstall

1. Stop the service:

   sudo systemctl stop claude-mem-worker
   sudo systemctl disable claude-mem-worker
   

2. Remove the plugin:

   rm -rf plugins/claude-mem/
   

3. Remove CLAUDE_MEM_URL from .env

4. Rebuild and restart NanoTars

Warning: Do NOT delete /root/.claude-mem/. This directory contains the shared memory database used by both the host and container agents. Removing the plugin only stops container access — the host's claude-mem continues to function independently.