← News & Insights

Blog

November 20, 2026

Attention is all you need, and all you have. Olyvr dreams to give it back.

The next job for AI is to give attention back to people.

Hasibul Haque, Syed Ishtiaque Ahmed, Yufei Wang, S M Mahbub Murshed, David Klubeck and Ryan Turner · Causal Dynamics Lab · November 2026

Your attention is the scarcest thing you bring to work, and it is under attack all day. Microsoft's study of its own Microsoft 365 users found that the average worker gets 117 emails a day and is interrupted every two minutes.

Of the people online at 6am, four in ten are already in their inbox. They are trying to work out the day's priorities.

The hard part of work is no longer doing it. It is deciding which work deserves you. An AI that waits for a prompt doesn't help with that. An empty chat box expects you to already know what is late and who is waiting.

A great executive assistant works the other way. They read everything while you're in meetings and sort it overnight. In the morning they hand you a short list: this is overdue, this person is waiting, start here.

Olyvr does that sorting while you're away. We call it Memory Dreaming.

Attention, for people

In 2017 a team of researchers, most of them at Google, published "Attention Is All You Need", the paper behind today's AI models. Its idea is simple. To understand a word, the model weighs every other word and attends to the few that matter.

That taught machines where to look. But the AI it made possible produces more messages than ever, and all of them ask for your attention.

Memory Dreaming applies the same idea to you. Instead of words, it weighs every message, ticket and document in your work. Instead of learned weights, it keeps salience: how much each person, project and promise matters to you today. Olyvr learns it from your own work, and you can correct it.

What makes Olyvr different

Background memory is an active research area. Databricks and Anthropic have both found that memory reorganised between sessions beats memory written only during the work. But their systems help an AI agent get better at its own tasks. Olyvr helps a person get their attention back, and that changes the design.

  • Whose memory. Yours, built from your mail, chat, calendar, tickets, documents and code, not an agent's own transcripts.

  • How it is organised. A map with you at the centre: people, projects, promises and the evidence for each, not a folder of notes.

  • When it speaks. First. Suggestions, your My Day list and draft replies are waiting when you arrive.

  • Where it lives. On your own computer, with any model, including a local one. Import needs no model at all: in our tests, 512,000 documents were built into memory in 106 seconds.

In our research, Skill Dreaming applies the same loop to procedures. It raised task success from 79% to 88% and from 83% to 92% on two difficulty tiers, with the model unchanged. It is research today, not yet a feature of the app.

A product manager's morning

Maya is a product manager at a fintech company, a week before a launch. When she opens Olyvr, suggestions under the question box already name her real people and open threads.

One says: "Draft a reply to Marcus in Compliance on his questions about payout limits." She didn't type it. Olyvr wrote it overnight.

[Image: Maya's home screen with suggestions]

My Day shows two promises that slipped while she was in launch reviews. Olyvr found them in a mail thread and a Slack channel she hadn't opened since Friday.

She presses Response on one, and a draft reply appears beside the list. It opens with a summary and asks who owns the open item, because that is how she works. Nothing is sent until she sends it.

Every answer also lists the sources it rests on, ranked by what matters to her. She checks one document instead of five threads.

[Image: My Day with the Response panel]

All names in this example are fictional.

How dreaming works

Dreaming is background work Olyvr does on your memory between sessions. It's named after the way the brain replays the day during sleep. It has four steps:

  1. Read. Olyvr imports your work into a private memory on your own computer.

  2. Dream. It reviews what happened and what your documents say, writes short digests, and tracks what keeps mattering.

  3. Know you. It updates a picture of you, your people and your projects that you can read and edit in Memory.

  4. Rank. It writes the next steps you see when you sit down.

Then it loops. What you click, finish or close feeds the next dream.

[Image: the four steps of Memory Dreaming]

Learning that stays honest

Background learning is only safe if it can't make things up. So Olyvr's software, not the model, enforces the rules:

  • Every learned line must quote its source word for word.

  • An observation needs two sources before it is shown.

  • Observations that new evidence contradicts are retired.

  • Anything you switch off stays off.

Rules you write yourself, such as "give all times in Eastern Time", always win over anything Olyvr learned.

Try it

Olyvr runs on Mac and Windows and is free during the beta. Connect your mail and Slack, let the first import finish, and see what Olyvr already knows about your day.

Download the beta

Sources