smart-memory

What This Skill Does

Smart Memory v3.1 acts as the cognitive backbone for your OpenClaw agent, transforming transient interactions into a structured, persistent knowledge base. Unlike simple chat history, this skill utilizes a transcript-first architecture that separates raw communication from derived insights. By maintaining a canonical SQLite store supported by a FastAPI engine, it ensures that your agent can recall specific preferences, task states, and complex identities with high precision. It features entity-aware retrieval, ensuring that when you ask for information, the agent doesn't just search for keywords, but understands the relationships between concepts, goals, and evidence. The system manages memory through explicit lanes—categorizing data into core and working memory—to ensure your agent stays focused on the immediate context while retaining long-term wisdom.

Installation

To integrate this skill, ensure you have the OpenClaw environment initialized. The core wrapper is located in skills/smart-memory-openclaw/. You can integrate it into your agent definition using the primary exports:

1. Import createSmartMemorySkill and createOpenClawHooks from the local skill directory.
2. Initialize the skill with your agent identity configuration.
3. Ensure your local environment satisfies the CPU-only PyTorch requirements for the underlying FastAPI engine.
4. Run the local backend server via server.py to enable the RESTful API endpoints.

Use Cases

• Long-term Project Continuity: Retain technical constraints and project goals across multi-day development sessions.
• Personalized Assistant Identity: Remember user preferences, communication styles, and past feedback to improve future interactions.
• Complex Decision Tracking: Audit the evidence behind an agent's past recommendations through revision-aware chains.
• Background Insight Generation: Allow the agent to process pending insights during idle periods to improve future response quality.

Example Prompts

1. "Look up my preferences for code formatting and ensure the current script follows them."
2. "What was the primary goal we established for the database migration project last week?"
3. "Summarize the decisions made regarding the API security audit based on our previous discussions."

Tips & Limitations

• Canonical Truth: Always remember that transcripts are the source of truth. If the memory state seems stale, trigger a rebuild via the /rebuild endpoint to re-derive insights from the transcript history.
• Retrieval Integrity: Avoid claiming prior context unless the memory_search tool returns high-relevance evidence.
• Performance: Keep the memory store performant by using the provided inspection endpoints to prune irrelevant memories.
• Safety: The skill performs file system writes to update the local SQLite store; ensure your runtime environment has appropriate write permissions in the /storage and /transcripts directories. The system will automatically queue failed commits, so keep an eye on your .memory_retry_queue.json.