What this diagram shows: OmegaClaw operates as a continuous cybernetic loop, not a request-response chatbot. Every cycle flows through four stages:
OBSERVE — The agent perceives new input (user messages, system events, time passing) and reconstructs context by querying both episodic memory (past interactions) and semantic memory (stored beliefs with truth values).
REASON — Using Non-Axiomatic Logic (NAL) and Probabilistic Logic Networks (PLN), the agent performs deduction, abduction, and evidence revision. Each derived conclusion carries a truth value (frequency, confidence) that degrades transparently through inference chains.
DECIDE — Candidate goals are ranked by priority, confidence, and alignment with long-term objectives. The agent genuinely decides what to do next. Low-confidence conclusions are deprioritized or flagged for more evidence.
ACT — The agent executes skills: shell commands, MeTTa expressions, file I/O, web search, memory storage, or messaging. Results feed back into OBSERVE, closing the loop.
The dashed box captures a key architectural insight: the LLM does not replace symbolic reasoning — it acts as an inference controller, steering which NAL/PLN chains to fire. The intelligence is in the steering, not raw computational speed.