persistent-memory

What This Skill Does

The persistent-memory skill provides a robust, three-layer memory architecture for OpenClaw agents. Unlike basic file-based recall, this skill creates a persistent, multi-faceted brain consisting of a human-readable Markdown repository, a semantic Vector Database powered by ChromaDB, and a relational Knowledge Graph built on NetworkX. By synchronizing these three layers, the agent gains the ability to not only store facts but to understand relationships between concepts and retrieve semantic context across long-term sessions. A core component of this skill is its automatic correction of OpenClaw's default search behaviors. By running the included configuration script, the agent gains visibility into critical directives like SOUL.md, AGENTS.md, and PROJECTS.md—files that are otherwise ignored by default configurations. This ensures your agent adheres to your defined behavioral rules and project constraints regardless of how long the interaction lasts.

Installation

The setup is designed for one-click deployment from your workspace root. Simply execute the command bash skills/persistent-memory/scripts/unified_setup.sh. This script handles the installation of all necessary Python dependencies, including ChromaDB, NetworkX, and sentence-transformers. It also performs a mandatory integration step that modifies your OpenClaw configuration to ensure critical workspace directives are indexed for every memory search. Once installed, ensure you run the indexer script after adding or modifying documentation in your memory directories to keep the vector database and knowledge graph aligned with your human-readable notes.

Use Cases

This skill is ideal for complex, long-running projects. Use it for maintaining a coherent "institutional memory" when working on multi-week software builds, tracking evolving business rules in a dynamic startup environment, or managing complex agent identities that require strict adherence to system prompts and behavioral guidelines across dozens of sessions. It effectively prevents "agent drift" where a model might forget a previously agreed-upon architectural decision.

Example Prompts

1. "Search my persistent memory for the architectural constraints we decided on regarding the database schema last week."
2. "Review the SOUL.md and MEMORY.md files, then provide a summary of our primary goals for this project to ensure we are aligned."
3. "Based on our past logs in the reference directory, how have we historically handled API rate limits for this specific provider?"

Tips & Limitations

To keep memory accurate, always run the indexer (vector_memory/indexer.py) immediately after editing your Markdown files. If you find the agent ignoring a specific instruction, verify your sync status using auto_retrieve.py --status to ensure your configuration is active. Keep your notes atomic and modular; the graph traversal performs best when concepts are well-linked in your source Markdown. Note that this skill requires local write access to your filesystem to maintain the index and logs.