lily-memory

What This Skill Does

Lily-memory serves as the primary long-term persistence layer for OpenClaw agents. Unlike transient chat sessions that disappear upon restart, this skill provides a hybrid database architecture utilizing SQLite FTS5 for rapid keyword-based lookups and Ollama-powered vector embeddings for nuanced, semantic understanding. It handles the lifecycle of agent knowledge automatically: capturing salient facts during dialogue, consolidating data to prevent bloat, and actively detecting when an agent is stuck in repetitive logic loops, allowing for a corrective nudge. By integrating directly into the OpenClaw plugin slot system, it enables agents to maintain continuity, remember user preferences, and reference past decisions across indefinite timescales.

Installation

1. Use the clawhub CLI to fetch the package: clawhub install openclaw/skills/skills/kevinodell/lily-memory.
2. Update your openclaw.json configuration file to include the lily-memory plugin within the plugins.entries object. Ensure the dbPath is set to a writable directory.
3. Verify that your Ollama server is running on the default port (11434) if you intend to utilize vector semantic search. If Ollama is unavailable, the plugin will gracefully fallback to basic keyword-only functionality.
4. Restart your gateway using openclaw gateway restart to initialize the database schemas.

Use Cases

• Project Management: Agents remember project status, deadlines, and team-specific terminology across long-running development sprints.
• Personal Assistance: Remember specific preferences, user-defined entities, and previous constraints established in earlier sessions.
• Knowledge Retrieval: Use the agent as a searchable repository for unstructured technical notes, internal documentation, and task-related decision logs.

Example Prompts

1. "Look up the status of the migration project from our discussion yesterday."
2. "Memory: The API endpoint for the production server is https://api.production.internal."
3. "Search my recent notes for any mention of the security policy update."

Tips & Limitations

• Vector Performance: While powerful, vector search depends on the speed of your Ollama embedding model. Ensure your hardware can handle model inferencing for a smooth experience.
• Storage: Since it uses SQLite, it is lightweight, but monitor your dbPath if you perform heavy, high-frequency logging to avoid massive file sizes.
• Validation: Stick to the entity allowlist in your configuration to maintain data integrity; auto-capture is regex-based and works best with explicit formatting.