Autonomous Cognitive Loop

A Home That Pays Attention · Deferred · home

Deferred

The catalog records this feature as deferred: it is not finished. The limits below say what is missing.

Perceive → Reason → Act → Observe → Reflect — runs on a configurable timer, only invokes the LLM when something novel happens.

The cognitive loop is the autonomous perception-reason-action tick that would run on a configurable interval (default 5-15 minutes). It gathers state from HA entities, Frigate events, scanner findings, the occupancy model, and behavior predictions. It uses cheap heuristics for routine checks and only invokes the LLM when something novel happens.

The loop runs in a background thread. PERCEIVE gathers state from all connected sources. REASON flushes pending events into the agent’s cognition queue. ACT routes desired actions through the AutonomyGate. OBSERVE checks whether actions had the intended effect. REFLECT logs the decision chain to the orchestration timeline.

Limits and invariants

The loop is interruptible — if the user sends a chat message while the loop is running, the chat takes priority. The action generation logic is currently a scaffold awaiting full implementation.

Where this lives

• halbert_core/halbert_core/home/cognitive_loop.py