How to use the Claude-Mem OpenClaw Plugin Setup agent skill

What does the openclaw skill make the agent do?

The skill walks you through setting up the claude-mem plugin on an OpenClaw gateway. When the setup is finished, your agents keep persistent memory across sessions through system prompt context injection. You can also turn on an optional real-time observation feed that streams new observations to a messaging channel.

The plugin works through four OpenClaw events: before_agent_start starts a claude-mem session, before_prompt_build injects the observation timeline into the system prompt (cached for 60 seconds), after_tool_call records each tool use such as Read, Write or Bash as an observation, and agent_end summarises the session and marks it complete.

What you need before you start

The setup assumes you already run an OpenClaw gateway. The plugin only targets that runtime, so it does nothing for Claude Code users who have no gateway. The installer and manual steps also expect the following tools on the machine.

• An OpenClaw gateway
• Bun, which the worker service uses (the installer checks for it)
• uv
• Node.js and npm, used to build the cloned claude-mem repo
• Git
• An AI provider API key if you pick a provider such as Gemini
• Optionally, a messaging channel for the observation feed

How do you install the openclaw skill?

Download the skill ZIP from this page and unzip it. For a personal setup, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. Claude Code will then find the skill.

If you use the Claude apps, you can upload the ZIP instead. Open Settings, go to Capabilities, then Skills, and upload the file there.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills

How do you install the claude-mem plugin on your gateway?

The skill points you to a one-line installer that handles dependency checks, plugin installation, memory slot configuration, AI provider setup, worker startup and the optional observation feed. It runs interactively unless you pass options. Only run the quick install if your environment allows piping a remote script into bash.

If you prefer to do it by hand, clone the claude-mem repository to a location your gateway can reach, run npm install and npm run build inside it, and then start the worker. Finally, add a claude-mem entry under plugins in your gateway configuration.

• Pre-select a provider and key with --provider=gemini --api-key=YOUR_KEY
• Skip prompts with --non-interactive (defaults to the Claude Max Plan and skips the observation feed)
• Upgrade an existing install, keeping settings, with --upgrade

How do you start the worker and check it is running?

The claude-mem worker is an HTTP service on port 37777. The plugin talks to it over HTTP, so it only needs to be reachable on localhost:37777. Check whether it is already running with the health endpoint. A response of {"status":"ok"} means it is up.

If Claude Code already has claude-mem installed, restart the worker from that installation. Otherwise, start it from your cloned repo. If the health check still fails, confirm bun is on your PATH, check npm run worker:status, and check whether another process is using port 37777 with lsof -i :37777.

• If you use Claude Code with claude-mem: cd ~/.claude/plugins/marketplaces/thedotmack then npm run worker:restart
• If you use the cloned repo: cd into it, then npm run worker:start
• Debug with npm run worker:status and npm run worker:logs (if available)

How do you configure the plugin and verify it works?

Add the claude-mem plugin to your gateway configuration. The project field scopes observations in the memory database, so use a different name for each gateway or use case, such as coding-bot. syncMemoryFile controls whether the timeline is injected into system prompts. Set it to false to stop injection while still recording observations. Restart the gateway, then check the logs for a line starting with [claude-mem] OpenClaw plugin loaded. You can also run /claude_mem_status in any OpenClaw chat. A feed status of disconnected is expected until you configure the observation feed.

To confirm recording, run an agent task and then open the worker viewer at http://localhost:37777 to see observations appear.

• project (default openclaw): scopes observations; use a unique name per gateway
• syncMemoryFile (default true): injects the timeline into system prompts
• syncMemoryFileExclude (default empty list): agent IDs that skip injection but are still recorded
• workerPort (default 37777): change only if you moved the worker

How do you ask an agent to use the skill?

Describe the task in plain language. Claude matches the request to the skill's description, so you do not need special syntax. The phrasings below are examples of the kind of request this skill covers, not a fixed list of trigger words.

Be specific about your gateway and channel. For instance, say which chat app you want the feed in, or whether the worker is already running on the machine.

• Set up claude-mem memory on my OpenClaw gateway
• Check whether the claude-mem worker on localhost:37777 is responding
• Walk me through adding a Telegram observation feed to my gateway
• My claude-mem worker is not starting, help me debug it
• Turn off context injection for the agent named reviewer

What are the safety notes for this skill?

This skill runs scripts and uses the network. The quick install downloads a script from install.cmem.ai and runs it with bash in one step. Read the script first if you want to inspect it, or use the manual route. The manual route clones a public GitHub repo and runs npm commands.

Be aware that tool uses are recorded as observations by default. If agent activity is sensitive, set syncMemoryFile to false or leave the observation feed off. The observation feed sends observation text to whichever chat channel you configure, so check who can read that channel.

The skill folder also includes a skills/agent-cost-report folder with Python scripts and a CHECKSUMS.txt file listing hashes for those files. Use it to check that the files are unchanged before you run anything.

• Quick install pipes a remote script into bash
• Observations are stored by default for every tool use
• Feed messages go to the chat channel you set, so check its members
• Verify bundled files against CHECKSUMS.txt

What tips help the setup go smoothly?

Use a unique project name for each gateway so memories do not mix. Run the health check before editing config, which tells you whether the worker is the problem or the plugin is. For the observation feed, find your chat ID first. In Telegram, group IDs are negative numbers. In Discord, enable Developer Mode to copy a channel ID. The feed only works for channel types that already run as plugins on your gateway. The repo's TESTING.md describes a Docker-based end-to-end test and a unit test suite if you want to check the plugin yourself.

• Check the health endpoint before changing config
• Give each gateway its own project name
• Telegram group chat IDs are negative numbers
• Discord: Settings → Advanced → Developer Mode, then right-click the channel and copy its ID