TurboQuant — Context Compression Plugin for OpenClaw

Inspired by Google's TurboQuant (ICLR 2026), this plugin brings the same hot/cold cache compression principle to OpenClaw at the application layer.

What it does

Every time you send a message to an AI, the entire conversation history goes along with it. The longer the session, the more tokens you're burning — and tokens cost money.

TurboQuant splits your conversation into two zones:

• Hot cache — the last N turns, kept verbatim (full fidelity)
• Cold cache — everything older, compressed to ~25% of original size

Net result: 40–70% fewer tokens sent on long sessions. Same quality, lower cost, faster responses.

Install

1. Clone into your OpenClaw extensions folder:

git clone https://github.com/Boehner/openclaw-turboquant ~/.openclaw/extensions/turboquant

2. Add to your openclaw.json:

{
  "plugins": {
    "allow": ["turboquant"],
    "load": {
      "paths": ["/path/to/.openclaw/extensions/turboquant"]
    },
    "entries": {
      "turboquant": {
        "enabled": true,
        "config": {
          "keepRecentTurns": 6,
          "compressionRatio": 0.25,
          "minTurnsBeforeCompression": 10
        }
      }
    }
  }
}

3. Restart the OpenClaw gateway.

Configuration

Option	Default	Description
enabled	true	Enable/disable the plugin
keepRecentTurns	6	Number of recent turns to keep uncompressed (hot cache)
compressionRatio	0.25	Target size for compressed turns (0.25 = 25% of original)
minTurnsBeforeCompression	10	Don't compress until conversation has this many turns

How it works

Uses extractive summarization — scores every sentence by information density (term frequency × position), keeps the highest-value sentences, drops the rest. No AI calls needed for compression — it's fast, deterministic, and free.

The algorithm mirrors TurboQuant's core insight: not all context is equally important. Recent turns matter most. Old turns can be compressed aggressively without hurting response quality.

Expected savings

On a 30-turn conversation:

• Without TurboQuant: ~3,200 tokens of history sent per request
• With TurboQuant: ~1,100 tokens (hot: 6 turns verbatim, cold: 24 turns at 25%)
• Savings: ~2,100 tokens per request (~66%)

License

MIT