memory-tiering

What This Skill Does

The memory-tiering skill provides a sophisticated framework for OpenClaw agents to manage their internal context window efficiently. By implementing a three-tiered architecture (HOT, WARM, and COLD), the skill ensures that the agent maintains peak performance while preserving essential history. The HOT tier acts as the immediate workspace for active tasks, keeping the context window uncluttered. The WARM tier serves as the agent's knowledge base for stable user preferences and configurations, while the COLD tier provides a long-term, summarized archive of past project milestones and historical decisions. This systematic approach allows the agent to handle large-scale projects without the latency and token-bloat common in long-running sessions.

Installation

To integrate this skill, run the following command within your terminal or agent console: clawhub install openclaw/skills/skills/sarielwang93/memory-tiering This will pull the necessary file structure to your local memory directory, establishing the three-tier file pathing required for operation.

Use Cases

This skill is ideal for complex, multi-day development projects where maintaining consistent context is difficult. Use it when you are transitioning between phases of a project, such as closing a feature branch and starting a new module, to ensure that the agent moves relevant data into its archives. It is also highly effective for users who want to define specific interaction styles or recurring system configurations that the agent should remember indefinitely without constantly re-processing them in every chat session.

Example Prompts

1. "Run memory tiering now to clean up after finishing the authentication module."
2. "整理记忆层级 - move the current project plan to cold storage and clear the active tasks."
3. "My timezone has changed to UTC+8; please update my preferences in WARM memory."

Tips & Limitations

To get the most out of memory-tiering, treat your COLD storage as a 'distilled knowledge base' rather than a dumping ground for raw logs; prioritize summaries over logs. For security, never store sensitive credentials in plain text; ensure your HOT memory configurations point to secure vaults. Note that this skill requires manual triggers for optimal control, though it does automatically run after /compact commands. Be mindful that aggressive pruning of the COLD tier can result in the loss of granular historical detail if summaries are not written effectively.