How OpenHome Is Scaling Personalized Agents To 80,000 Users With Memorable.
OpenHome Superagent
TASK COMPLETIONProcedural memory against flat retrieval on the same tasks.
OpenHome Superagent
WASTED TOOL CALLSCounted across every eval run.
OpenHome builds an open smart speaker platform. Each device exposes around 150 keyword-triggered abilities that route into one agent, the superagent. The hardware has little onboard storage.
150 Abilities, Each Starting From Zero.
Every ability re-derived its approach on each call. The more steps a request had, the longer the person waited in silence. A procedure written into the prompt was paid for on every call, by every user, whether the run needed it or not.
Built For A Device With Limited Storage.
Most memory products store a conversation in the cloud. OpenHome needed procedures that fit on the device, for an agent that acts on requests.
Before Memorable
FLAT RETRIEVALSession logs sized for a chat history.
Sized for a server, far above the storage budget of a speaker.
Preferences carry across sessions. The steps of a task do not.
With Memorable
PROCEDURAL MEMORYThe steps that finished the task and the result that proved it.
Fits the storage budget of a speaker.
The same steps run again. No new plan is reasoned out.
Routing Daemon, Trace Flywheel, Superagent.
Routing Memory Daemon
Sits on the device between the wake word and the ability it triggers. It checks for a stored procedure before the superagent starts reasoning.
100% Task Completion, Up From 70%.
| Metric | Before (Flat Retrieval) | After (Procedural Memory) | How it was measured |
|---|---|---|---|
| Task completion | 70% | 100% | Same tasks, flat retrieval against procedural memory |
| Wasted tool calls | Present | 0 | Every eval run |
| Live demo completion | 50% | 100% | 8 live tasks, memory off then on |