Learning, privacy and human control

How a persona learns

Within a conversation it may adjust pacing, explanation and questions. That does not alter its shared identity. Shared learning begins with aggregated signals: recurring questions, missing knowledge, misunderstandings and cultural shifts. Personal details are excluded by default; any future collection needs an explicit privacy and retention procedure.

A repeated claim is a research lead, not evidence of truth. Coordinated users, duplicate reports and malicious sources can manipulate apparent consensus. Research proposed changes against independent evidence where appropriate and keep external text separate from operating instructions.

The proposed loop is: signals → research → change proposal → Gauntlet → human decision → versioned release → monitoring. There is currently no shared conversation collection or automatic learning service.

Human approval is mandatory

Initial recommendation: weekly review during development; monthly review once stable. This is a meeting cadence, not an automatic release schedule. No scheduler has been created.

The review packet contains the current and proposed versions, exact changes, supporting and conflicting evidence, intended effect, Gauntlet results, remaining failures, privacy implications and rollback target. A named human approves, rejects or requests changes. Record the decision, date and exact reviewed snapshot. Editing that snapshot invalidates approval. Agents may research and propose patches; they may not approve their own work or publish persona revisions.

Critical failures may justify immediately disabling a feature or reverting to an already approved version, with an incident record and human notification. They do not justify inventing and deploying a new identity automatically. Approved updates must not introduce new facts outside their reviewed scope.

Today, Pre is enabled as a local voice review candidate; Marriott Hotels remains a research draft and cannot create paid voice sessions. Guinness remains an illustrative prototype. This eligibility flag is a local guard, not an authenticated production approval system. Production will need reviewer identity, immutable approval records and access control before automated proposals are connected.

Temporary interaction logs

Temporary conversation logging is a proposed evaluation feature, not operating today. The CEM dashboard contains invented transcripts only. For brand-hosted experiences, plan client-controlled records with permissions, encryption, retention and deletion. Sponsored Agent log export is unconfirmed; a private cloud does not make a transcript anonymous. Agree what is collected, for what purpose, who can read it, retention duration and deletion procedure. Prefer minimal text and session diagnostics over audio. Give people clear disclosure and appropriate controls. Sensitive disclosures must not become sales leads or shared cross-brand knowledge.

Dashboard insights should distinguish evidence from inference: unanswered questions, grounded factual errors, character issues, interruptions, latency and usage cost. Aggregated learnings may propose a shared patch; a brand manager reviews changes to their distinctive character. No conversation silently rewrites shared DNA. Cross-brand learning and confidential data use require explicit agreement.