ARIA — Adaptive Relational Intelligence Agent

Open the app. Say nothing. ARIA speaks first. "Hey, I noticed you've said you're behind on the migration three times this week — always in the afternoon. Want to talk through what's actually blocking you?" Nobody asked her to. She worked it out from your last conversations. Knows your day. Remembers your world. Speaks up before you ask. ARIA is a voice-first companion that works like a chief of staff on the phone. Your calendar doesn't know how the last meeting went, your news feed doesn't know your client, and your AI forgets you when the tab closes. ARIA connects all of it. Five pillars • Memory — people, facts and promises, kept across weeks • Schedule — Google Calendar by voice, tied to who's in each meeting • World — live news that touches your work • Proactive — tracks what you promised, flags conflicts, opens next time with one insight • Emotional continuity — your mood shapes her tone; the trend shapes what she notices She talks like a friend, follows your language, and knows the time where you are. Built on AssemblyAI. One Voice Agent API WebSocket runs the whole call — Universal-3 Pro speech-to-text, turn detection, the LLM, the voice and 8 JSON-schema tools — with mid-session keyterms, mood-swapped prompts and live word-by-word captions. The LLM Gateway is her memory: every turn it extracts promises, facts and mood, and after the call it writes the insight ARIA opens with next time. Hard problems solved: non-English replies re-voiced via Google TTS, a safety net for skipped tool calls, and a rate-limit circuit breaker. 51 automated tests. Honest edges: single user; emotion is read from text. Most assistants wait to be asked. ARIA speaks up first.