Token Optimizer Skill for OpenClaw

An optimization suite for OpenClaw agents to prevent token leaks and context bloat.

Features

• 🔒 Cron Isolation — Run background tasks without polluting your main context
• 🔍 Local RAG — Semantic search over memory files instead of loading everything
• 🔄 Reset & Summarize — Protocol for context consolidation when hitting limits
• 📜 Transcript Indexing — Search through old session transcripts
• ⚡ Hybrid Search — Combine vector and keyword search for better recall

Installation

Via ClawHub (coming soon)

openclaw skill install token-optimizer

Manual Installation

1. Clone this repository into your skills folder:

cd ~/.openclaw/workspace/skills/
git clone https://github.com/D4kooo/Openclaw-Token-memory-optimizer/tree/main

2. The skill will be automatically detected by OpenClaw.

Quick Start

1. Isolate Background Tasks

In your openclaw.json, set sessionTarget: "isolated" for cron jobs:

{
  "cron": {
    "jobs": [
      {
        "name": "My Background Task",
        "schedule": { "kind": "every", "everyMs": 1800000 },
        "sessionTarget": "isolated",
        "payload": {
          "kind": "agentTurn",
          "message": "Do the thing. Use message tool if human needs to know."
        }
      }
    ]
  }
}

2. Enable Local RAG

Configure semantic search for your memory files:

{
  "memorySearch": {
    "embedding": {
      "provider": "local",
      "model": "hf:second-state/All-MiniLM-L6-v2-Embedding-GGUF"
    },
    "store": "sqlite"
  }
}

3. Use the Reset Protocol

When context exceeds 100k tokens:

1. Ask your agent to summarize the session
2. Update MEMORY.md with important facts
3. Run openclaw gateway restart

Configuration Reference

Setting	Description	Default
memorySearch.embedding.provider	Embedding provider (local, openai)	—
memorySearch.embedding.model	Model for embeddings	—
memorySearch.store	Storage backend (sqlite, memory)	memory
memorySearch.paths	Paths to index	["memory/", "MEMORY.md"]
cron.jobs[].sessionTarget	Session type (main, isolated)	main

Why This Matters

Long-running OpenClaw sessions accumulate tokens from:

• Heartbeat checks
• Background task results
• File reads and tool outputs
• Conversation history

Without optimization, you'll hit context limits and experience:

• Slower responses
• Higher API costs
• Lost context when truncation kicks in

This skill teaches your agent to stay lean.

Contributing

PRs welcome! Areas we'd love help with:

• [ ] Native hybrid search implementation
• [ ] Automatic context monitoring
• [ ] Smart transcript archiving
• [ ] Cost tracking integration

Credits

• shAde — Original concept
• Clément — Implementation

License

MIT — Use freely, credit appreciated.

---

Part of the OpenClaw ecosystem. 🦦