By 28 June, the central companion loop joined three previously separate ideas. Typed conversation could route between deterministic responses, a limited fallback, and a guarded Apple on-device provider. Spoken input could be started deliberately, reviewed as a transcript, and submitted through the same pipeline. Approved local memory could begin to inform eligible low-risk replies.
The implementation kept each boundary narrow. Speech never auto-submitted. Unsafe or high-stakes responses remained text-only. Ordinary conversation did not become memory automatically.
An explicit remember command created a proposal, and only human approval allowed a distilled low-risk record to persist.
This was a technical integration milestone rather than product validation. Physical-device audio, real Apple model behavior, conversational quality, memory wording, and accessibility all remained pending human review.
← Back to Cirilo