Description:
MeTTaClaw is a compact, operational agentic AI system implemented entirely in MeTTa, running on the PeTTa compiler. In roughly 200 lines of core code, it delivers a fully transparent and inspectable agent featuring embedding-based long-term memory, tool use (web search, file modification, shell access), and skill acquisition — all within the symbolic MeTTa framework.
Unlike opaque LLM-based agents, MeTTaClaw's entire decision-making process is visible as Atomspace operations. Its embedding-indexed Atomspace enables semantic retrieval over past interactions, while timestamped command traces provide a complete audit trail. The agent learns new skills and declarative knowledge through memory items incrementally, building competence over time.
MeTTaClaw serves as both a practical agentic tool and a foundation for more sophisticated cognitive architectures. The community can contribute code, use cases, and teach the running agent new skills via memory and tool extensions.