KaiCalls

Calling a business back should not mean explaining your problem all over again. KaiCalls is a business phone service with a secretary built in. We added Muse to help that secretary find the details that matter from earlier conversations. Our demo follows four real human test calls about printing from a phone. The caller corrects iPad to Android, then adds that the phone is a Pixel. The recordings show both a missed detail and later recall. KaiCalls' existing memory path automatically extracts facts after calls and supplies saved history to subsequent calls; those four calls demonstrate that foundation, with GPT-4.1 handling the conversation. Meta's Muse Wonder adds a separate reasoning layer. For explicit topic recall, Muse selects relevant fact IDs; application code checks the caller, business and source records, then returns the stored text with its source-call references. In an earlier staged phone test, Muse selected two manually seeded printer facts in 2.288 seconds, and the provider transcript records the secretary using them. Separately, we replayed the first human call through Muse's structured extraction path: the output preserved the Android correction, printer problem and Wi-Fi/app uncertainties, validated against the schema, and took 3.440 seconds at a reconciled cost of $0.013008. Muse runs through OpenRouter. It does not replace the speaking model or run during call pickup. The extraction replay did not write live memories. The public repository reproduces the memory engine and bench; private phone adapters remain outside that excerpt. The result is a working caller-memory foundation and a measured Muse integration, with a clear path toward a business that remembers its customers.