Under the hood

A plain-language look at how Nora works — and how she's designed to get better over time — without losing the one promise that matters: she only acts with your OK, and always from her own identity, never as you.

The simple version

Nora is built from three simple ideas:

• A scheduled helper — like a night-shift assistant, Nora can wake on a schedule, do one small job, and leave you a note. She never runs off on her own forever.
• A fact-checking team — for bigger jobs (like research), several helpers work in parallel and a skeptic double-checks the facts before Nora shows you anything.
• Learning while idle — in quiet moments, Nora reviews what's happened and writes herself useful notes, so she's sharper next time — kept private, inside your own boundary.

And underneath all of it, the same rule never changes: propose → you approve → Nora acts.

How a request flows

flowchart TD
    U["You forward an email to Nora"] --> N{"Nora"}
    N -->|"drafts a reply, in your voice"| D["A draft — you send it"]
    N -->|"researches, fact-checked"| R["A sourced answer"]
    N -->|"proposes a meeting or a file share"| P["A proposal"]
    P --> A{"You approve? (reply AG)"}
    A -->|"yes"| ACT["Nora acts — from her own labelled identity, never as you"]
    A -->|"no reply yet"| F["A gentle reminder, a few days later"]
    F --> A
    N -. "in quiet moments" .-> L["Nora learns — notes kept inside your data boundary"]
    L -.-> N

Everything that reaches another person passes through the approve step. The draft is always yours to send; the few things Nora does (invite someone, share a file, book a meeting) happen only after your explicit go-ahead.

A closer look

Fact-checked research (rolling out)

When you ask Nora to research something, instead of a single quick lookup she'll put a small team on it: helpers dig into different angles at the same time, a verifier tries to disprove shaky claims, and Nora hands you an answer with its sources. It takes a little longer, but you can trust what comes back.

Friendly follow-ups (rolling out)

If Nora proposed something — say, a meeting with a partner — and you got busy and never approved it, she'll gently remind you a few days later: "Still want me to set this up?" She re-surfaces the proposal; she never sends or books anything on her own. It's careful by design: she won't nag twice for the same thing, and she can't act by accident.

Learning while idle — "Nora dreams" (rolling out — voice first)

In quiet periods, Nora consolidates what she's seen into useful memory: who the recurring people are, how you like your drafts written, helpful context on ongoing threads. Next time, her drafts and summaries start smarter — and she does this without retraining any AI model, simply by keeping good notes.

Because this touches your real correspondence, it carries the strongest safeguards:

• It stays inside your boundary. Memory for a cloud-tier mailbox stays in the cloud tier; for a device-only setup, nothing leaves the device. Memory never crosses that line.
• A safety check on every note. Before anything is saved, it's checked — so a malicious email can't trick Nora into "remembering" something false or harmful.
• Private to each person, and off by default — Nora only ever learns from what she's allowed to see.

The constant

However Nora grows, three things stay true:

1. Bounded — she works in short, finite bursts, never an endless loop.
2. Safe by default — when unsure, she does the cautious thing (and asks).
3. You're in control — nothing reaches another person without your explicit approval, and Nora never sends as you.

That's the same promise as the rest of Nora: she drafts, you send — and she acts only with your OK.