acc-error-memory

What This Skill Does

The acc-error-memory skill acts as an AI agent's internal monitor, analogous to the anterior cingulate cortex in the human brain. It is designed to track error patterns, escalate recurring mistakes, and learn effective mitigations over time. By providing a structured framework for logging failures—such as tone mismatches, hallucinations, or context loss—this skill ensures that your agent evolves through persistent memory across sessions. It employs a sophisticated severity ranking system where errors are classified as normal, warning, or critical based on frequency, eventually triggering automated maintenance to keep your agent's performance optimized.

Installation

To integrate this skill, navigate to your OpenClaw workspace and execute the installation script. Open your terminal and run:

cd ~/.openclaw/workspace/skills/anterior-cingulate-memory ./install.sh --with-cron

This process initializes the necessary acc-state.json file for tracking and configures a cron job to run pattern analysis three times daily (4 AM, 12 PM, and 8 PM). Ensure your ACC_MODELS environment variable is configured to define the LLM screening hierarchy, allowing the system to use models like Claude, OpenAI, or Ollama for error classification.

Use Cases

This skill is essential for agents operating in long-running projects where consistency is paramount. It is ideal for debugging persistent logic errors in coding agents, maintaining brand-consistent tone in customer support bots, and preventing recurrent factual inaccuracies in research agents. By tracking "something's off" moments, developers can identify the specific scenarios where their agent consistently fails and apply targeted mitigations.

Example Prompts

1. "OpenClaw, I noticed you keep misremembering the project framework version. Please log this as a recurring factual error for the acc-error-memory skill and suggest a mitigation strategy."
2. "Run the load-state script to show me which error patterns have been escalated to critical severity in the last 24 hours."
3. "Resolve the 'tone_mismatch' error pattern, as I've adjusted the system prompt to handle that context correctly moving forward."

Tips & Limitations

• Model Selection: Use high-reasoning models (like Sonnet or GPT-4o) in your ACC_MODELS list for the best classification accuracy. Using lower-tier models for screening may result in false positives.
• Resolution: Patterns clear automatically after 30 days of inactivity, but manual intervention is recommended if you have implemented a fix to prevent the agent from over-correcting.
• Persistence: Ensure that your environment variable for ACC_MODELS is exported in your shell profile so that background cron jobs can access the models successfully.