memento

What This Skill Does

Memento serves as the long-term memory engine for OpenClaw agents, transforming fleeting interactions into a structured, persistent knowledge graph stored locally in SQLite. By monitoring conversation streams, Memento systematically captures key information, extracts actionable facts, and recalls relevant context before every AI turn. It ensures agents remain consistent, remembering user preferences, previously discussed decisions, and specific task requirements without needing constant reminders. The architecture is designed with a privacy-first philosophy, utilizing local storage and supporting local LLM inference via Ollama to ensure data never leaves your environment unless explicitly configured for external API processing.

Installation

To integrate Memento into your environment, use the OpenClaw CLI to fetch the package from the repository: clawhub install openclaw/skills/skills/braibaud/memento. Once installed, perform a gateway restart to initialize the database schema. Ensure your openclaw.json is updated with your desired configuration, specifically setting autoExtract to true and selecting your preferred extractionModel. For advanced context retrieval, download the BGE-M3 embedding model to the local ~/.node-llama-cpp/models directory to enable high-performance semantic search capabilities alongside standard keyword-based FTS5 lookup.

Use Cases

• Project Management: Track ongoing task status, blockers, and project decisions across multiple long-running conversations.
• Personal Assistance: Manage complex user preferences, such as communication styles, preferred scheduling hours, or recurring meeting requirements.
• Agentic Knowledge Management: Allow multiple specialized agents to share a unified pool of factual knowledge while maintaining boundaries for sensitive data (private vs. secret).
• Context Preservation: Resume work across distinct sessions without needing to re-summarize past events or explain technical constraints repeatedly.

Example Prompts

1. "Based on what we discussed last week about the backend architecture, what was the primary reason we chose PostgreSQL over MongoDB?"
2. "Update my profile: I now prefer that all project summaries be delivered in a concise bulleted list rather than long paragraphs."
3. "Summarize all the action items assigned to me from the last three meetings we had regarding the Q4 launch."

Tips & Limitations

• Privacy Guardrails: Always use local models like Ollama if you are handling sensitive or proprietary data to ensure total air-gapped security.
• Model Tuning: Adjust minTurnsForExtraction to prevent unnecessary API calls for short, irrelevant chatter.
• Knowledge Boundaries: Be mindful that while crossAgentRecall enhances collaboration, sensitive data should be clearly labeled to prevent leakage between agent scopes.
• Performance: Semantic search (BGE-M3) is significantly more accurate than standard keyword search but requires more local compute overhead.