Responsible: Ben Goertzel
Description:
HyperClaw is a cognitive orchestration layer that extends MeTTaClaw's agentic foundations to coordinate heterogeneous AI modules — multiple LLMs, local code executors, symbolic reasoners, and neural subsystems — through an attention-metaprotocol governing what each module attends to and when.
Designed by Ben Goertzel, HyperClaw embeds MeTTaClaw as its core agent loop and adds three key abstractions:
- Typed Context Frames: Structured MeTTa representations of multi-module cognitive state — goals, hypotheses, experimental history, module capabilities, and results. These are inspectable, queryable, and version-controlled as Atomspace objects.
- Two-Timescale Attention-Metaprotocol: MeTTa rewrite rules operating at two timescales — a fast-loop (seconds) for task execution and module selection, and a slow-loop (minutes–hours) for strategic steering and mode transitions.
- Module-Space Abstraction: Every external module — LLM endpoint, Python executor, data store, sensor stream — is wrapped as a uniform Hyperon Space supporting match, bind, and rewrite operations.
HyperClaw deepens the inter-module communication fabric of a full Hyperon/PRIMUS cognitive architecture, where PLN, MOSES, ECAN, and predictive-coding networks all interoperate over the shared Atomspace.