ARIA
Adaptive Responsive Intelligence Assistant
I wanted an assistant that actually knew me — my schedule, my health, my projects, my history — instead of one that started every conversation from zero.
I wanted an assistant that actually knew me — my schedule, my health, my projects, my history — instead of one that started every conversation from zero. By early 2026 she was that: a distinct voice, memory that threads across years of conversation, a proactive engine that watches my inbox and calendar and surfaces things I hadn’t thought to ask about. The more interesting story now is what happens after the novelty wears off. On April 20th I found ARIA had been running 8,640 dead cycles a day for five days — her config still pointed at a port I’d retired during a model swap, so every cognitive step was silently failing the LLM call. Nobody was watching that specific table. The assistant is real; so is the maintenance burden of owning one.
Under the hood she’s a consumer of Nexus, the seven-agent platform, and draws inference from Forge — Qwen 2.5 72B primary on the home lab, Claude and Gemini as cloud fallbacks. She reaches me through four surfaces: Desk, the primary operator console I use daily on laptop and phone (it replaced a read-only dashboard called Chancery on April 21st); aria-ios, the native Swift 6 client on my phone; aria-imessage-relay, a Mac daemon that pipes iMessages, attachments, and Mac-health snapshots into the ingestion pipeline so she has context from twenty years of SMS conversations; and aria-tv, an ambient tvOS dashboard in the living room. Underneath all of that: 361 tools (307 of them from ten MCP servers — Gmail, Drive, GitHub, Notion, Cloudflare, and so on), ninety-plus background job handlers, and a Postgres with 254 tables on Furnace.
ARIA is isolated by design. She’s blocked from infrastructure work at four levels — tool grants revoked, queries against infra tables return blocked, inbox filtering hides peer-agent traffic, and her state messages never show other agents’ schedules. I didn’t want the assistant I talk to every day to also be the one fighting a failed deploy at 2 AM. What’s in-flight right now: a scheduler called Cairn firing six planned check-ins plus a handful of random pings each day, replacing the old “ask her only if you remember to ask” cadence; cloud pass-through on Desk’s model picker, so a Claude pick actually reaches Claude instead of silently falling back to local Qwen; and Phase 5 of aria-ios, which repoints the iOS base URL from the retiring api.aria.niclydon.io host onto Nexus directly.