Skip to content

How it works

An intelligence that keeps the thread.

Most AI answers a message and forgets it. Nova runs a loop — it perceives your work, remembers it across sessions, reasons about what you're doing, acts across your desktop, and learns from the result. The context survives, so it already knows where you left off.

In plain English

What Nova is.

Nova is software that runs on your computer and holds the context of your work over time.

Every assistant you’ve used lives inside one window and forgets the moment you close it. You re-explain the project, the files, the decisions — again and again. That re-explaining is the tax you pay for tools with no memory.

Nova removes it. It keeps a persistent, private record of what you’re working on, on your device, and works across your whole desktop rather than inside a single app. Because it remembers, it can pick up where you left off, notice what connects, and do the small things without being briefed from scratch.

The difference isn’t a smarter answer. It’s continuity — an intelligence that understands your work as it accumulates, not one conversation at a time.

The loop

Seven stages, always turning.

This is how Nova is built to work. Each stage feeds the next, and the last feeds back into the first — so every pass is a little sharper than the one before.

The loop, end to end
Perceive
Remember
Reason
Plan
Act
Verify
Learn
Learning feeds back into perception — the loop closes, and the next pass starts further ahead.

Stage by stage

How information moves through the system.

Each stage receives something and produces something for the next. Follow the receives and produces lines and you can trace a piece of work all the way around the loop.

01

Perception

receivesscreen · files · apps · system state
producesa current picture of your work

Nova reads what is actually on your machine — the window you're in, the files a project touches, the app you just opened. It sees the work, not a description of it.

02

Memory

receivesthe current picture + your intent
producesthe records that bear on the task

It ties what it sees to what it already knows: this project, these people, the decision you made last week. Relevant records are retrieved before anything is done.

03

Reasoning

receivespicture + retrieved context
producesan understood goal, or a question

It works out what you're trying to accomplish and whether it has enough to proceed — or whether it should ask. Understanding comes before action.

04

Planning

receivesthe understood goal
producesa plan of reversible steps

A request becomes a small sequence of concrete, reversible steps. Nothing sweeping happens in one opaque move.

05

Action

receivesthe plan, one step
producesa real change in your apps

It operates real applications the way you would — opening, editing, moving, drafting — one step at a time.

06

Verification

receivesthe result + the goal
producesa pass, or a correction

It checks the result against the goal and its own constraints, and catches and corrects its mistakes rather than pressing on.

07

Learning

receiveswhat worked + your corrections
producesupdated memory

What worked, what you corrected, and what you prefer are written back to memory — so the next turn of the loop starts further ahead.

produces: updated memory → receives: the next picture · the loop closes

The loop in motion

One real pass, start to finish.

The same seven stages, following a single task — resuming a coding project after two weeks away.

This is illustrative. The loop it shows works today; stronger reasoning and deeper software-development workflows are what we’re still expanding.

  1. 01

    Perception

    You reopen a project you left two weeks ago. Nova sees the repository, the editor, and the branch you're on.
  2. 02

    Memory

    It retrieves what you were building — the payment flow — the reason for the last change, and the constraint a teammate asked you to keep.
  3. 03

    Reasoning

    It works out that you're resuming the refund path, and that the edge case you flagged is still open. It has enough to help, so it doesn't need to ask.
  4. 04

    Planning

    It proposes three reversible steps: re-run the failing test, isolate the branch that breaks, and draft the fix for your review.
  5. 05

    Action

    With your go-ahead it runs the test and prepares the change — one step at a time, each visible.
  6. 06

    Verification

    It checks the fix against the constraint and the test, catches that one case still fails, and says so instead of declaring success.
  7. 07

    Learning

    Your correction and the working approach are written back — so next time it starts from the fix, not the blank page.

Desktop intelligence

Why it isn't a chatbot.

Other tools are excellent at what they do. The difference is architectural — where the intelligence lives, and whether it remembers. No criticism intended; just the shape of each.

A chatbot
ScopeOne chat window
MemoryForgets between chats
ContinuityStarts from zero
Browser AI
ScopeThe current tab
MemorySession-bound
ContinuityEnds with the tab
An IDE assistant
ScopeOne editor
MemoryThe open repository
ContinuityScoped to the file tree
A computer-use agent
ScopeA fresh sandbox
MemoryNone between runs
ContinuityWakes up blank
Desktop automation
ScopePre-written scripts
MemoryNo understanding
ContinuityRepeats fixed steps
Nova
ScopeYour whole desktop
MemoryPersistent, on-device
ContinuityCarries across sessions

Nova’s foundation is built: memory, perception, file understanding, and desktop control work today. The loop above runs; stronger reasoning, better verification, and deeper development workflows are what we’re strengthening next.

Questions

The loop, answered plainly.

Is Nova watching my whole screen, all the time?
Only within the scope you grant, and it pauses to ask before anything irreversible. What you keep out of scope stays invisible to it. The full boundary is on privacy & security.
How is this different from an agent that just clicks around my screen?
Two things: memory and verification. Nova retrieves what it already knows before it acts, and it checks its own results against the goal instead of pressing on. A blank agent does neither.
What happens when it gets something wrong?
The plan is a sequence of small, reversible steps, and verification is built to catch mistakes rather than hide them. You keep the undo, and the correction is remembered so the same error doesn't repeat.
Do I prompt it, or does it act on its own?
You're in the loop. Nova proposes and acts on your behalf, not instead of you — and it asks before anything it can't cleanly undo.
How much of this loop works today?
The foundation is built: memory, perception, file understanding, and desktop control all work today. Stronger reasoning and better verification are what we’re strengthening now, and deeper software-development workflows are in active development — see the roadmap for the honest state of each.

See it do real work.

Nova is in private preview. Request access, and we’ll write when it’s your turn.

Request access