crabpath

What This Skill Does

CrabPath is a high-performance memory graph engine designed to act as the long-term cognitive layer for AI agents. Unlike standard vector databases that simply store embeddings, CrabPath maintains a structured, traversable graph of information, enabling sophisticated retrieval, real-time correction, and autonomous learning. The core engine is built with a 'pure' design philosophy—it requires no external network calls or hidden system dependencies, ensuring total control over your agent's knowledge representation. It operates using a canonical state.json format, making it highly portable and auditable. Users can define custom embedding logic and LLM callbacks, allowing the agent to integrate seamlessly with services like OpenAI for intelligent routing while keeping the graph logic isolated and deterministic.

Installation

To install this skill, use the ClawHub CLI tool: clawhub install openclaw/skills/skills/jonathangu/crabpath

Use Cases

• Long-term Agent Memory: Maintain state across multiple sessions to create persistent, evolving AI personalities that remember user preferences and past project contexts.
• Knowledge Graph Management: Organize complex, hierarchical information where nodes represent specific facts or documents, and edges represent logical relationships.
• Real-time Correction Flows: Implement a 'learning loop' where the agent receives feedback, corrects its internal graph nodes, and prevents future hallucinations or errors.
• Autonomous Research: Store and traverse large datasets, summarizing information dynamically to provide high-context, efficient responses without exceeding token limits.

Example Prompts

1. "Initialize a new CrabPath workspace at ./project_docs and index all text files using the OpenAI embedding provider."
2. "Query the graph for project requirements using the state file 'memory.json' and apply a beam search width of 8 for deeper contextual extraction."
3. "Inject a correction into the current graph state: the project architecture is microservices-based, not monolith, and update the related nodes accordingly."

Tips & Limitations

• Traversal Limits: Monitor your beam_width and max_hops. Setting these too high can drastically increase response latency. The default max_context_chars of 20,000 is a safe starting point to prevent context window overflows.
• State Management: Since the skill relies on a state.json file, ensure your backup strategy accounts for this file to prevent data loss.
• Pure Logic: Because the core is pure, avoid hard-coding environmental secrets into your callback functions. Use configuration injection for API keys to maintain a clean security boundary.
• Health Checks: Regularly run crabpath doctor and crabpath measure_health to prune stagnant nodes and ensure your graph remains performant and lean.