A private, voice-first thinking partner that quietly builds a structured library of your life and work while you talk to it — so it can hold you accountable, do your research, and remember everything you'd rather not.
You talk to it; it doesn't just answer — it files. Every project, person, decision, commitment and deadline you mention is captured into a knowledge base it keeps for you, in a shared format you never have to see or organize. Over time it becomes a second brain that knows your context and gets sharper the more you use it. You don't set it to a role — it becomes one by listening.
One tool, many roles — set entirely by what you talk about: exercise coach · trip guide · business operator · job-hunt cadence · thesis organizer · habit & mood check-ins · story-world bible · research desk.
A private, per-person space on our own servers. A background scribe captures what matters, a scheduler turns intentions into dated check-ins, a steward reconciles and corrects — you never tag or file anything. Knowledge compounds, can be corrected, and segments into projects. The AI model and voice engine are swappable; your library is the constant.
Architecture. Three tiers: a static gate (Cloudflare Pages), an always-on persistence hub plus a thin LLM brain shim — both containerized on our own Linux servers (Docker on Fly) — and a voice layer. Nothing is locked to a black-box SaaS.
Multi-tenant, individual-first. Each user is a hard-isolated tenant: their own database and knowledge files in a per-user slot, resolved per request. The model could federate tenants into shared team spaces — deliberately switched off today.
Model-agnostic. The brain is a shim. Anthropic's Claude now (managed API — fast to ship, cheap to run for proof-of-concept); pointing it at a self-hosted open-source model is a config change, not a rewrite.
Voice, swappable. Gradium.ai speech-to-text / text-to-speech today; the target is Unmute by Kyutai running our own model, fully self-hosted — managed now for ease and cost, self-hosted when it earns it.
Memory substrate (OKF). Atomic, versioned knowledge files are the single source of truth, indexed in SQLite for fast recall — segmented by project, correctable, supersedable. The shared format is invisible to you.
Async agents over one graph. A scribe captures, a scheduler dates work and books check-ins, a researcher looks things up in the background, a steward reconciles / merges / corrects, a summarizer closes the loop — all coordinating through OKF so the conversation stays fast and heavy work runs off the hot path.
Always-on. The hub also works between conversations: finalizing dropped sessions, reconciling memory, firing "it's time" reminders (in-app or Telegram). The intelligence lives in the persistent substrate — the servers — not just the chat window.
Private by default · individual-first · structure you never have to manage.
🎧 Find a quiet spot — headphones or a noise-cancelling mic make voice clearest. The agent handles the notes; you just talk.
This account is using a temporary password — choose your own to continue.