Atlas: Private On-Device AI

* Atlas looks ahead. Before upcoming meetings with people Atlas knows, it surfaces relevant open reminders as a heads-up. It can also optionally draft a short 2–4 step plan for what's coming up. Plans are suggestions only and never acted on automatically. Heads-ups are fully on-device; suggested plans are off by default and require background enrichment. * Atlas learns from you. Thumbs down feedback can include corrections that Atlas distills into lasting lessons and applies to future answers. Atlas can also learn durable preferences, goals, habits, and constraints from confirmed “remember…” requests and, when background enrichment is enabled, from conversations. Everything is clearly marked, reviewable, editable, and deletable under Settings -> Models & Behavior. Stored entirely on-device. * More reliable background learning. Relationship and durable-fact extraction now uses structured, schema-constrained generation, making the knowledge graph and personal self-model less likely to fail due to malformed model output. * Updated on-device AI engine. Atlas now runs on LiteRT-LM v0.15.0. Existing downloaded models continue working without re-downloads. A light repetition penalty also reduces rare runaway repetition from smaller models. * Streaming answers. Chat replies now appear word-by-word as they're generated instead of all at once, while tool use and confirmations continue to work as before. * Actions actually happen. Requests such as “remind me…” and “add lunch with Sarah” now reliably execute the requested action after explicit confirmation, rather than merely claiming they were completed. Atlas can also show your reminders and track when Atlas-created reminders are completed. Actions and outcomes are reviewable under Settings → Diagnostics → Actions. * Better date awareness. Atlas now knows today's date and can reason about dates such as “three days from now,” with more reliable answers to questions like “what day is it?” * Prepare for tomorrow. The Briefing tab brings together tomorrow's calendar, related people and notes, and open reminders into a private on-device preparation view. * Answer feedback. Every answer now has thumbs up/down buttons, with thumbgs down allowing a quick correction. Feedback is stored only on-device and can be reviewed or deleted under Settings → Diagnostics → Answer Feedback. * Atlas just answers. Everyday questions — recipes, recommendations, how-tos, ideas, and similar requests — now receive direct answers instead of being treated as knowledge-base searches. Atlas can personalize these answers using what it knows about you. “I searched and found nothing” is reserved for questions about your own notes and documents. * A richer knowledge graph. Atlas now connects life information as well as work relationships, including recommendations, visits, and things you like, want, or own. Near-duplicate names and shortened forms can be conservatively merged into the same person or place, with every merge reversible. * Smarter answers about people and places. When you ask about someone or somewhere in your saved knowledge, Atlas automatically brings relevant knowledge-graph connections into the answer. * Saved web pages become searchable. .html files placed in a watched folder are converted into readable article text while excluding menus, scripts, and styling. Processing happens entirely on-device; Atlas never fetches the page over the network. * Showcase everywhere. The Demo Showcase and Super Demo are now available on both iPhone and iPad. * Clearer model downloads. Interrupted or timed-out transcription model downloads now show the actual connection error instead of “model not found,” with a Retry button that resumes the download rather than starting over. * Cleaner model loading. Removed the duplicate “Preparing model…” indicator so only a single loading status is shown.