The meeting was with the chief executive, and it mattered. Sophie Leroy was a consultant then, juggling a growing number of projects and switching between them all day.
When it ended, she told her team they had forgotten to ask one question.
“They looked at me funny and said, ‘No, we did, we asked it,’” she recalled in a 2024 profile by her university. For a few minutes, she realized, her brain had been on another project entirely. The proof was on her notepad: a little to-do list with items from her other projects, written during the meeting.
The tab that stays open
Switch tasks before you’ve let go of the last one, and part of your mind stays behind.
Leroy went into research, and in a 2009 paper titled “Why is it so hard to do my work?”, gave the problem its name: attention residue. Across two experiments, she reported, people found it hard to pull their attention away from a task they hadn’t finished, and they did worse on the task that came next.
An unfinished task behaves like a browser tab left open in the background. You’re looking at a new page, but the old one is still running and still drawing on the machine.
The same paper had a twist. Finishing the first task wasn’t enough to let it go. In her results, the people who disengaged were those who had finished under time pressure, and that pressure contributed to better performance on the next task.
Most work, though, doesn’t end on a deadline. It gets interrupted.
Worry about getting back
In 2018, Leroy and Theresa Glomb of the University of Minnesota published four studies on what interruptions do to the task that does the interrupting. Her university’s 2024 profile describes the designs.
People were pulled off one task and onto another. Some were told they’d have plenty of time to finish the first task later; others were told they wouldn’t. Those expecting a squeeze carried more residue into the new task. They processed information less carefully, missed errors and were less likely to find the best answer to a decision problem.
Then the pair tested a remedy they called a “ready-to-resume” plan. Before switching, you jot down where you were and what you planned to do next. Less than a minute, Leroy said, may be enough. In a field study and a lab study, people who made the plan switched their attention more fully and did better on the interrupting task. The idea borrows from research on implementation intentions: a plan for when and how you’ll come back lets the mind set the task aside for now.
How far this reaches is still being tested. In a 2025 study, Ward van Zoonen and Yuri Scharp noted that the plan had mostly been tested in laboratory or cross-sectional settings, often with students. In their own field test, 50 employees rated their attention residue over four workdays; on days with more intrusions they reported more residue and more exhaustion, and writing ready-to-resume plans reduced the residue. Unexpectedly, the employees who wrote the plans also reported more exhaustion overall, even though residue weighed less on their exhaustion from day to day. Those were self-ratings from a small group.
The bug you leave half-fixed
Imagine you’re an engineer halfway through a payment bug when you’re pulled into a budget call. The release is tomorrow, and you know you’ll have to rush the fix afterward.
That combination, an interruption with a squeeze waiting on the other side, is the one that left more residue in Leroy and Glomb’s studies. So before you join, you spend forty seconds writing in the ticket: stopped at the retry loop, next step is to check the timeout, one open question for the payments team. Then you close the editor. In their studies, that kind of note helped people give the interrupting task more of their attention. It won’t fix the bug. But tomorrow starts from the note instead of from memory.
Leftovers in the context window
Language models have a version of residue you can actually see. Whatever a session keeps of its history takes up the context window, the text the model reads before each answer. Some tools summarize older turns when the window fills, but what remains is read again with every answer.
In a July 2025 report, researchers at Chroma, which makes database software for AI applications, tested 18 models, including GPT-4.1, Claude 4, Gemini 2.5 and Qwen3, on questions about long chat histories. Each question came two ways: as the full history, about 113,000 tokens of mostly irrelevant conversation, or as a focused prompt of about 300 tokens holding only the parts needed to answer. Across all models, the focused prompts scored significantly higher. The full prompts were both longer and full of other material, so the test doesn’t separate length from clutter.
Claude Code’s documentation gives the matching advice. It warns against the “kitchen sink session,” where you start one task, ask about something unrelated, then go back, leaving the context “full of irrelevant information.” The fix it gives is /clear between unrelated tasks.
The difference from a person is that a model’s residue sits on the page, where you can delete it. Before clearing, ask the agent to write a ready-to-resume note: what it changed, where it stopped, what comes next. Start the next session from that note. To see whether it matters for your work, run the same small task a few times in a long mixed session and a few times in a fresh one, and judge each result against the task’s own test. One better run is a lead, not proof.
Leroy’s notepad from that meeting held exactly that kind of list. She had just written it at the wrong time.