memory

What This Skill Does

The Memory skill provides OpenClaw agents with a persistent, structured, and protocol-driven long-term memory system. Unlike standard context windows that clear at the end of a session, this skill implements a three-layered architecture covering protocol, capture, recall, and maintenance. By leveraging SESSION-STATE.md for hot context and RECENT_CONTEXT.md for rolling highlights, the skill allows agents to retain specific user preferences, project decisions, and historical data across multiple restarts. It solves the common issue of context degradation by enforcing a 'Write-Ahead Log' policy where agents update local files before processing complex queries, ensuring critical information is preserved regardless of conversation length.

Installation

To integrate this system into your OpenClaw environment, execute the following command in your terminal: clawhub install openclaw/skills/skills/rosepuppy/memory-complete

Once installed, initialize your workspace by copying the provided templates: cp skills/memory/references/SESSION-STATE.md ./ cp skills/memory/references/RECENT_CONTEXT.md ./

Finally, update your AGENTS.md to include the Memory Protocol instructions and add the capture hooks to your HEARTBEAT.md configuration to enable automatic state management.

Use Cases

• Project Continuity: Maintain complex project requirements, technical debt trackers, and architectural decisions over days or weeks.
• Personalization: Remember user preferences, coding styles, and feedback loops without needing to restate them in every new session.
• Context Recovery: Seamlessly resume work after an agent restart or context compaction by loading the 'hot' session state from disk.
• Knowledge Management: Organize conversation history into actionable facts and actionable TODOs through the periodic consolidation utility.

Example Prompts

1. "Based on our previous architectural discussions, what were the two main reasons we decided against using a relational database for the cache layer?"
2. "I'm updating my project preferences; please save the fact that I now prefer using Tailwind CSS for all front-end styling in this repository."
3. "Recall our notes from last week regarding the authentication module and summarize the pending tasks that haven't been marked as complete yet."

Tips & Limitations

• Maintenance: Run consolidate.py regularly to prevent the accumulation of redundant facts and keep the recall search speed optimized.
• Scope: This system is best suited for factual data and decision logs. Avoid storing extremely large binary blobs or excessive raw text; instead, rely on the extraction logic within capture.py to keep memory footprint low.
• Protocol Adherence: The effectiveness of this system relies entirely on the agent following the mandatory protocol in AGENTS.md. Ensure your system prompt is strictly enforced for best results.