An assistant should remember.
An AI that forgets who you are every morning is not an assistant. It is a stranger wearing a familiar voice. Tanya includes a seven-level memory architecture — from working memory to biographical — and is being tested for durable continuity across deployed surfaces.
An assistant should say "I don't know."
Hallucinations are not bugs. They are the predictable output of systems optimised to sound fluent rather than to be calibrated. Tanya includes an optional grounding path that can check claims against available sources. When that path runs, low confidence can be surfaced; it is not guaranteed on every response.
An assistant should evolve.
The pace of AI research is inhuman. No model trained six months ago is current. Tanya is being built to watch the ecosystem and help you adopt new, vetted capabilities — so you are less likely to wake up to a stale tool.
An assistant should be yours.
Inspect and self-host the MIT-licensed source on infrastructure you control. A managed service is a future target and is not currently offered for sale. Your deployment, your responsibility.
An assistant should be built in the open.
No hidden harness. No marketing leaps. Reviewed releases ship behind public issues, public plans, and honest changelogs. If Tanya fails, you will see it fail. If it breaks new ground, you will see exactly how.
An assistant should act when it matters.
Apple Intelligence parity is the floor, not the ceiling. The differential we are building toward is agency — the ability to surface unprompted when something deserves your attention. Manifestations are scarce, earned, and always under your structural consent.
This is the minimum bar for a cognitive partner in 2026.
Anything less is just a chatbot.