---

name: memory-setup description: Configure and validate OpenClaw memory recall for persistent context. Use when enabling memory_search/memory_get, fixing poor memory recall, or setting up MEMORY.md + memory/*.md workflows in an OpenClaw workspace.

Memory Setup (OpenClaw)

Set up durable memory for OpenClaw so the agent can recall prior decisions, preferences, and todos.

1) Prepare workspace files

In workspace root, keep:

• MEMORY.md (curated long-term memory)
• memory/YYYY-MM-DD.md (daily notes)

Optional structure:

• memory/projects/
• memory/system/
• memory/groups/

2) Enable memory search in OpenClaw config

Configure under agents.defaults.memorySearch (not top-level memorySearch).

Example:

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "enabled": true,
        "provider": "local",
        "includeDefaultMemory": true,
        "maxResults": 20,
        "minScore": 0.3
      }
    }
  }
}

Notes:

• includeDefaultMemory: true indexes MEMORY.md + memory/**/*.md.
• Providers: local, openai, gemini, voyage, mistral.
• If using remote provider, set corresponding API key (env var or memorySearch.remote.apiKey).

3) Restart and verify

• Restart gateway after config changes.
• Verify with:
  • openclaw status
  • openclaw memory status (if available in your CLI build)

4) Test recall behavior

Ask a question about past context, then verify the agent:

1. Runs memory_search
2. Uses memory_get for precise lines when needed
3. Cites source path/lines when useful

5) Troubleshooting

memory_search unavailable

• Ensure agents.defaults.memorySearch.enabled = true.
• Ensure policy allows memory tools.
• Restart gateway.

low-quality matches

• Lower minScore (e.g. 0.2) to broaden hits.
• Raise maxResults (e.g. 30).
• Write more specific notes in MEMORY.md and daily logs.

local provider issues

• Confirm local model path/settings if configured.
• If needed, switch provider to remote and set API key.

6) Recommended operating rule

Before answering questions about prior work, decisions, dates, people, preferences, or todos:

1. memory_search first
2. memory_get second (only needed lines)
3. Say you checked memory if confidence is still low

This keeps responses grounded and auditable.