OpenHome: faster, cheaper voice agents

Voice is unforgiving about latency. There's no spinner to hide behind, no progress bar to watch. Every repeat task — reminders, flight check-ins, automated routines — was being rebuilt from scratch: the model re-read its instructions, re-planned its steps, and re-paid the token bill.

That cost lands twice. Users wait through step-by-step model calls, and OpenHome pays for tokens spent re-deriving procedures the agent had already figured out.

Memorable stores how things get done, not just what happened. The system watches successful agent runs and converts them into reusable workflows — described by objective rather than implementation, so they survive model and tool changes.

On a repeat task, the agent matches the stored workflow and executes it in milliseconds. Procedure documentation moves out of the prompt entirely, so every call gets leaner. Moving procedures out of the prompt is where the cost gets paid back — and it gets paid back twice: latency and tokens.

Live testing ran eight real tasks, each run twice. With procedural memory enabled, all eight completed — near-perfect completion, steady across repeat runs. Stored workflows executed in milliseconds versus step-by-step model processing, and token costs dropped through leaner API calls.

Multi-step workflows — reminders, flight bookings, automated routines — now execute from stored procedures across OpenHome's fleet of 100K live agents.