OpenClaw Skill

Trunkate AI partners directly with OpenClaw to offer an officially maintained Skill. This integration provides workspace-based prompt injection with automated lifecycle hooks for silent, “invisible” memory management.

Installation

The recommended method is via ClawdHub, which automatically manages dependencies and hook registration.


# Install the Trunkate AI skill
clawdhub install trunkate-ai
 
# Set your API key
export TRUNKATE_API_KEY="tk_live_1234..."

Lifecycle Hooks

Trunkate operates primarily via automated triggers. Configure these in your .openclaw/config.json to ensure memory is managed as a background utility.

Proactive Pruning (PreRequest)

Prunes history before every LLM call to maintain constant context density.


{
  "hooks": {
    "PreRequest": [
      {
        "type": "command",
        "command": "python3 hooks/openclaw/pre_request.py"
      }
    ]
  }
}

Emergency Recovery (OnError)

Triggers on context failures (e.g., 400/429 errors) to perform emergency history stabilization.


{
  "hooks": {
    "OnError": [
      {
        "type": "command",
        "command": "python3 scripts/error_detector.py"
      }
    ]
  }
}

Preservation Hierarchy

To protect critical project requirements, Trunkate follows a strict semantic preservation hierarchy:

Content Type	Priority	Strategy
Atomic Blocks	Critical	100% Verbatim; wrap in `[KEEP]` tags.
System Rules	Critical	100% Verbatim; never sent to API.
Active Goal	High	Full fidelity retention of user intent.
Recent Turns	High	Verbatim retention of last 3-5 turns.
Logic Blocks	Medium	Semantic compression via Trunkate API.
Large Logs	Low	Summary transformation (error extraction).

Environment Variables

Variable	Requirement	Purpose
`TRUNKATE_API_KEY`	Required	Authentication for the API.
`TRUNKATE_AUTO_BUDGET`	Optional	Target for history (e.g., `2000` or `"20%"`).
`TRUNKATE_DEBUG`	Optional	Enable verbose logging of hook execution.

Note: For manual optimization or focus pivoting, use the standard openclaw bridge: openclaw "Optimize my current context for [task]".