halcyon / Reading

Notes, not notifications

6 minute read research attention

— halcyon team

The default shape of an AI assistant, as of 2026, is a stream. It arrives in chunks. It nudges. It badges the app icon. It writes a follow-up question before you have read the answer. The interface is built on the assumption that the user's attention is a resource the assistant should claim, not a resource the assistant should protect.

There is a quieter shape. Notes, not notifications. An agent that finishes a piece of work and leaves it on the desk. No buzz, no badge, no streaming monologue. You will find it when you come back to it. That is the entire interaction model.

This is not an aesthetic preference. It is what the literature says works.

Sophie Leroy's Why is it so hard to do my work? (Organizational Behavior and Human Decision Processes, 2009) is the canonical paper on attention residue — the part of cognition that stays with task A when you move to task B. Leroy's experiments showed that residue is heaviest when the prior task was left in suspension: open, unbounded, unfinished. Every notification is, by design, a task left in suspension. You did not finish reading; you broke off mid-thought. The residue compounds across the day.

The cost is not theoretical. Gloria Mark's The Cost of Interrupted Work (CHI 2008) followed thirty-six knowledge workers through their days. After a single interruption, returning to the original task at the same depth took, on average, twenty-three minutes and fifteen seconds. The interruption did not need to be long. The recovery did.

A more recent strand of research has put the smartphone on the bench. In The hidden cost of a smartphone (PLOS ONE, 2022), participants who heard a smartphone notification while performing an attention task responded more slowly and showed larger N2 ERP responses — the electrophysiological signature of recruited cognitive control. The notification did not have to be answered to cost something. It only had to arrive. A 2025 PNAS Nexus paper, Blocking mobile internet on smartphones improves sustained attention, found measurable gains in sustained attention and well-being when notifications and mobile internet were turned off for a sustained period.

Aggregate these and you get a small, consistent picture. Interruptions are expensive. The cost is paid in attention residue, in stress, in slower responses, in elevated cognitive control. The cost is paid whether or not the user "feels" interrupted, because some of the cost is below the threshold of self-report.

Now look at what most AI assistants do. They are, structurally, notification engines. They stream. They suggest. They prompt for follow-up. They badge. They are, in the language of the literature above, optimised to generate residue.

We took a different default in halcyon. There is no streaming token, because the streaming token is a UI primitive that exists to keep you looking. There is no autocomplete in the composer, because autocomplete is a low-grade interruption masquerading as help. There is no follow-up nudge after a draft arrives, because the silence after a draft is part of the work. The agent finishes, lays the result down, and stops. You read it on your schedule.

The interaction we are reaching for is the one you have with a notebook. A notebook does not ping. It does not stream. It does not nudge. You go to it when you have something to do, and when you do, it has held everything you left in it. The notebook is the calm baseline against which a generation of "AI assistants" has measured itself, and the comparison has not been flattering.

A note, in our usage, is a complete artefact. A draft. A plan. A short script. A research summary with citations attached. The agent produces it, and it sits there. A notification is the opposite: a fragment whose only purpose is to demand the next move. The first is something to read. The second is something to dismiss. We do not think those are the same shape.

The strongest version of the argument is this. An assistant that interrupts you to deliver intelligence is, by the available cognitive science, withdrawing intelligence from the work. If the agent's contribution is gated by your ability to think clearly after reading it, then any cost it imposes on your clarity is netted against its own output. Notifications are a negative term in that equation.

We would rather ship a quieter product than a louder one that, by the measurements above, is paying for its own demos out of the user's working memory.

Sources