TurboQuant Optimizer

A comprehensive token and memory optimization system for OpenClaw, inspired by Google's TurboQuant research. Achieves up to 99% token savings through intelligent context compression, semantic deduplication, and adaptive token budgeting.

Description

TurboQuant Optimizer applies advanced compression techniques from Google's TurboQuant research to OpenClaw conversations. It operates at three levels:

1. Session Level: Intelligent context compression and summarization
2. Message Level: Semantic deduplication and content optimization
3. Token Level: Adaptive token budgeting and smart truncation

Key Innovations:

• Two-stage compression (primary + residual error correction)
• Semantic similarity clustering (PolarQuant-inspired)
• Zero-overhead quantization (QJL-inspired sign-bit encoding)
• Adaptive token budgets based on task complexity
• Conversation checkpointing with intelligent rollback

Installation

openclaw skills install turboquant-optimizer

Configuration

Add to ~/.openclaw/openclaw.json:

{
  "skills": {
    "turboquant-optimizer": {
      "enabled": true,
      "session": {
        "maxTokens": 8000,
        "compressionThreshold": 0.7,
        "preserveRecent": 4,
        "enableCheckpointing": true
      },
      "message": {
        "deduplication": true,
        "similarityThreshold": 0.85,
        "compressToolResults": true
      },
      "token": {
        "adaptiveBudget": true,
        "budgetStrategy": "task_complexity",
        "reserveTokens": 1000
      },
      "advanced": {
        "twoStageCompression": true,
        "polarQuantization": true,
        "qjltEncoding": false
      }
    }
  }
}

Usage

Automatic Mode

Once enabled, optimization happens transparently:

// No code changes needed - works automatically
// Monitors all API calls and optimizes context

CLI Commands

# Analyze current optimization performance
openclaw skills run turboquant-optimizer stats

# Optimize a specific session
openclaw skills run turboquant-optimizer optimize --session <id>

# Run benchmarks
openclaw skills run turboquant-optimizer benchmark

# Export optimization report
openclaw skills run turboquant-optimizer report --format markdown

Programmatic API

const { TurboQuantOptimizer } = require('turboquant-optimizer');

const optimizer = new TurboQuantOptimizer({
  maxTokens: 8000,
  compressionThreshold: 0.7
});

// Optimize messages
const optimized = await optimizer.optimize(messages);

// Get detailed statistics
const stats = optimizer.getDetailedStats();
console.log(`Token efficiency: ${stats.efficiencyScore}/100`);

How It Works

Two-Stage Compression (TurboQuant-Inspired)

Stage 1 - Primary Compression (PolarQuant-style):

• Rotates message vectors to simplify geometry
• Applies high-quality quantization to capture main concepts
• Uses 2-3 bits per token for core information