Give your OpenClaw agent persistent memory

OpenClaw is a powerful agent framework, but every session starts from zero. Your agent forgets the codebase knowledge it built up, the mistakes it learned to avoid, and the preferences you expressed. GreenCube fixes this by sitting between OpenClaw and your LLM provider as a transparent proxy that remembers everything.

Setup takes 5 minutes. No code changes to your agents.

Setup

1 Install GreenCube

curl -fsSL https://greencube.world/install.sh | bash

Windows: irm greencube.world/install.ps1 | iex

2 Open your OpenClaw config

Edit ~/.openclaw/openclaw.json and add GreenCube as a provider:

{
  "models": {
    "mode": "merge",
    "providers": {
      "greencube": {
        "baseUrl": "http://localhost:9000/v1",
        "apiKey": "your-openai-key-here",
        "api": "openai-completions",
        "models": [
          {
            "id": "gpt-4o",
            "name": "gpt-4o",
            "reasoning": false,
            "input": ["text"],
            "contextWindow": 128000,
            "maxTokens": 16384
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "greencube/gpt-4o"
      },
      "models": {
        "greencube/gpt-4o": {
          "alias": "gpt-4o"
        }
      }
    }
  }
}

Replace your-openai-key-here with your actual API key. Change the model if you use something other than gpt-4o.

3 Restart OpenClaw

openclaw daemon restart

Every request now routes through GreenCube before reaching OpenAI. Responses come back unchanged — your agents don't know anything is different.

4 Check what your agent learned

curl localhost:9000/brain

After a few tasks, you'll see facts it extracted, domains it tracks competence in, and a log of everything it noticed. Run this anytime.

What your OpenClaw agent gets

Persistent memory — facts from past sessions get injected into future tasks automatically

Self-reflection — after each task, the agent extracts what it learned and stores it

Self-verification — the agent reviews its own output and flags quality issues

Competence tracking — knows it's 87% on Python and 43% on CSS, adjusts caution accordingly

Mistake prevention — catches known error patterns before they reach you, retries automatically

Memory decay — useful knowledge gets reinforced, noise fades over time

How it works

GreenCube runs a local proxy on localhost:9000. When OpenClaw sends a request to greencube/gpt-4o, GreenCube receives it, injects relevant knowledge from past tasks, forwards it to OpenAI, then analyzes the response in the background. Knowledge extraction, self-verification, and competence tracking all happen asynchronously — the response time is unchanged.

Everything runs on your machine. No cloud storage. No telemetry. Your API key goes directly to OpenAI, just routed through a local process first. GreenCube is free, open source, and MIT licensed.

Get GreenCube