How to use the MemOS Local Memory Guide agent skill

What agent skills are and what this one does

An agent skill is a folder with a SKILL.md file that tells an AI agent when and how to use a set of tools. This skill does not add memory by itself; it describes how the agent should use the MemOS memory tools that are already provided by the MemOS Local plugin.

Once installed, the agent can search your past conversations, check your stated preferences and earlier decisions, and write its own shorter search query when automatic recall finds nothing. It can also share memories with other agents in the same workspace, share with a team server when you agree, and pull reusable how-to guides saved from earlier tasks.

• Search past chats, preferences and decisions with memory_search
• Read the full original text of a hit with memory_get
• Follow saved how-to guides from earlier tasks with skill_get
• Share or unshare memories with local agents or a team server

Install the skill from the ZIP

Download the ZIP for this skill from the page where you found it, then unzip it. The skill is delivered as a folder containing SKILL.md. Place that folder in one of these locations:

For a personal install, put it in ~/.claude/skills/. For a single project, put it in the project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal: ~/.claude/skills/
• Project only: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Check the prerequisites before you rely on it

The skill only works when the MemOS Local plugin is installed in an OpenClaw workspace. Without it, the tools named in the skill will not exist and the guide will not work.

The skill's metadata lists these requirements: Node.js and npm, curl or wget, better-sqlite3 (installed via npm), the environment variables OPENCLAW_STATE_DIR and OPENCLAW_CONFIG_PATH, and the configuration file ~/.openclaw/openclaw.json. A team server configuration is needed only for team sharing and is optional.

• Node.js and npm
• curl or wget
• better-sqlite3, installed via npm
• OPENCLAW_STATE_DIR and OPENCLAW_CONFIG_PATH environment variables
• ~/.openclaw/openclaw.json
• MemOS Local plugin and an OpenClaw workspace
• Team server configuration (optional, for team sharing)

Trigger the skill with natural requests

You do not need a special command. The skill is meant to be used whenever you refer to past chats, your own preferences or history, or when the agent needs prior context to answer. Ask in plain language.

If the agent's automatic recall returns nothing, which can happen with a long or unclear question, the skill tells the agent to write a short, focused query and run memory_search itself. Shorter, specific wording usually helps the agent find the right memory.

• "What did we decide about the database schema last week?"
• "Remind me what settings I said I prefer for code formatting."
• "Check our earlier chats on the deploy error and what fixed it."
• "Find how we set up the staging workspace before, and follow that approach."

Understand the sharing actions and their limits

The skill separates two sharing planes. Local agent sharing is visible only to agents in the same OpenClaw workspace. Team sharing is visible to teammates through the configured team server.

Per the skill's instructions, sharing with other local agents may happen without asking. Sharing to a team server (target hub or both) should happen only after you explicitly consent, and the agent should not share to the hub silently. Memory may contain personal or confidential content, so review what the agent proposes to share. Use memory_unshare to remove a memory from local agent sharing, the team, or both.

• memory_search: default scope is local; group or all includes team memories
• memory_write_public: creates a new local shared note and does not publish to the team server
• memory_share: target agents (default), hub, or both
• memory_unshare: removes sharing, default target all

Quick check that it is working

Ask the agent about something you discussed in an earlier session in the same workspace. If it returns relevant excerpts, memory search is working. If it says nothing is found, rephrase with fewer, more specific words, or confirm that the MemOS Local plugin and environment variables are set up.

If a result looks relevant but is cut short, the skill says to fetch the full text of that chunk with memory_get, which returns up to 4000 characters by default and 12000 at most.