agents-remember

Install For OpenClaw

OpenClaw uses an agent workspace with workspace files and skill folders. MCP server definitions live in the OpenClaw config registry under mcp.servers.

Official references:

Root Starter Package

The repository includes an OpenClaw starter package at .openclaw/. Copy that folder to your shared workspace root, then render the copied package. The .openclaw/render-starter script is a convenience: with a single --repo list such as --repo my-app shared-lib, it infers the workspace root from the copied .openclaw/ folder, fills path and repository placeholders, validates that each requested repository exists, and leaves .openclaw/openclaw.merge.json ready to merge into ~/.openclaw/openclaw.json before you restart the gateway or start OpenClaw. If you prefer not to run the renderer, make those same replacements by hand and verify that no placeholder tokens remain.

The package contains:

If you already have an OpenClaw workspace, copy the contents of .openclaw/workspace/ into that workspace instead of changing agents.defaults.workspace.

After the restart, invoke:

c-13-install-and-onboard

That skill runs or verifies runtime_install() and then handles memory, onboarding, and providers.

Workspace Instructions

Put the Agents Remember instruction in the OpenClaw workspace AGENTS.md, pointing at the actual coordination runtime path OpenClaw can read:

# Workspace Agent Instructions

Read and follow `/path/to/ar-coordination/AGENTS.md` before working in any target project.
Treat these rules as workspace instructions!

@/path/to/ar-coordination/AGENTS.md

OpenClaw workspaces may contain other standing instruction files. Keep Agents Remember focused on repository memory and task workflow rules; do not put secrets in workspace docs.

Do not overwrite a target repository’s root AGENTS.md just to wire Agents Remember; root AGENTS.md is project-specific. The starter package uses the OpenClaw agent workspace file instead.

OpenClaw has internal hooks and typed plugin hooks, but the documented file-based hook surface is for Gateway automation and lifecycle events, not a minimal project-local MCP starter. The starter package therefore uses the documented workspace AGENTS.md bootstrap file as the first-action authority.

OpenClaw loads bootstrap files from the configured agent workspace. If the workspace path or bootstrap files change while the gateway is running, restart the gateway.

MCP

OpenClaw stores third-party MCP server definitions under mcp.servers in ~/.openclaw/openclaw.json. The starter package includes this merge template:

{
  "mcp": {
    "servers": {
      "agents-remember": {
        "command": "uvx",
        "args": [
          "--refresh-package",
          "agents-remember-mcp",
          "agents-remember-mcp@latest",
          "--config",
          "<PATH/TO/YOUR/PROJECTS_FOLDER>/.openclaw/mcp/agents-remember-settings.json"
        ]
      }
    }
  }
}

After merging, use these OpenClaw commands to inspect the saved registry:

openclaw mcp status --verbose
openclaw mcp doctor

openclaw mcp probe agents-remember opens a live connection and is useful when you want proof that the MCP server starts correctly.

Runtime And Skills

After the restart, the copied c-13-install-and-onboard skill runs or verifies:

runtime_install()

OpenClaw loads bundled skills, managed/local skills under ~/.openclaw/skills, and workspace skills under <workspace>/skills, with workspace skills taking precedence. The starter package uses .openclaw/workspace/skills.

Do not run skills_install() for first-run setup. The copied starter package already provides one flat folder per skill under .openclaw/workspace/skills/. skills_install() remains available for manual maintenance and non-package installs.

You can inspect OpenClaw skill visibility with:

openclaw skills list
openclaw skills check

Long-running Turns

Deep reasoning and large onboarding waves usually fit inside OpenClaw’s current agent runtime timeout. If a local or remote model stops streaming long enough to hit the provider idle watchdog, prefer a provider-scoped timeout:

{
  models: {
    providers: {
      "<provider-id>": {
        timeoutSeconds: 300
      },
    },
  },
}

Use agents.defaults.timeoutSeconds only when the entire embedded agent run needs a different runtime cap. Restart the gateway after config changes that do not hot-apply:

openclaw gateway restart
This site is open source. Improve this page.