Why You Cannot Categorize Time From Memory

Sam Corcos logged 17,784 hours of work across five years as CEO of Levels, in 15-minute blocks, into no more than ten categories (First Round Review, 26 March 2025, updated 5 May 2025). That record is why he can tell you that recruiting, which felt like about 20% of the job, was about 3%, and that strategy, which felt like it ate the week, was 5% (First Round Review, 9 August 2021).

If he had reconstructed those years from memory, or from a calendar of meetings, he would have been wrong about the mix. I want to talk you out of reconstructing the week from memory. The practice is contemporaneous capture, then a small category set.

How Corcos Actually Logged the Hours

Every 15 minutes gets a category. There are at most ten. Anything that does not fit goes into Operations. He keeps a rigorous calendar and back-fills how he used the day, the same day (First Round Review, 9 August 2021). Five years in, he was still tracking every 15-minute block because the data kept him honest about tradeoffs (First Round Review, 26 March 2025).

The calendar is doing two jobs, and neither is Cal Newport's intention-blocking. The first is a constraint: he killed his to-do list because items on a list are untethered from time, and he answers "can you have this by Friday?" with the open hours he actually has. The second is a log: the filled day, back-filled that evening, is what he later categorizes. Newport assigns hours before they happen. Corcos uses the calendar to stop overcommitting, then treats the filled day as data.

He started because a tracker showed more than three hours a day on social media, against a guess of about 20 minutes. Information processing, mostly email, ran 3-4 hours a day, and in August 2020 averaged 34 hours a week (First Round Review, 9 August 2021). None of that looks like a meeting. A calendar will not save it.

Recruiting Felt Like 20 Percent. The Log Said 3.

Across the first two years he put 7,922 hours into Levels. The felt mix and the logged mix disagreed, in both directions.

Recruiting gave him anxiety: searching, getting rejected, asking for introductions. He would have guessed at least 20% of his time. The log said about 3%, and even a month that felt like all recruiting still topped out at 14% (First Round Review, 9 August 2021).

Strategy went the other way. He had written hundreds of pages of internal memos, and it still felt like a large share of the job. It was 5% in the two-year analysis, and 5% again across five years: 924 hours out of 17,784 (First Round Review, 26 March 2025). He did not publish a numeric expected share for strategy. He said it felt like much more than it was.

His own gloss is the finding: work that is demanding or emotionally taxing feels like it took more time than it did, and work that is easy or energizing feels like it took less (First Round Review, 9 August 2021). A reconstructed week is a mood board, not a category mix.

Memory Keeps the Peak and Throws Away the Duration

This is not a claim that psychologists followed founders around with clipboards. Redelmeier and Kahneman asked patients to rate pain during colonoscopy or lithotripsy, then to rate the procedure as a whole. Retrospective totals tracked the peak and the end, not how long the procedure lasted (Redelmeier & Kahneman, Pain02994-6), 1996). Fredrickson and Kahneman had already shown the same pattern with film clips: stretching an episode did little to change the global retrospective evaluation (Fredrickson & Kahneman, JPSP, 1993). They called it duration neglect.

A week of knowledge work is not a colonoscopy. The analogy is only this: when you later name the mix, you overweight the intense episodes and underweight the long, unremarkable stretches. Corcos's recruiting and strategy gap is what that looks like in a real job.

Duration recall fails even when the question is simpler than "what share was recruiting?" Ernala and colleagues compared ten common Facebook time questions against server logs for 49,934 people in 15 countries. The best survey question still only reached r = 0.42 against the logs. The open "how many hours a day" question produced reports of about 4 hours against 1.3 hours actual, a 3.2-hour daily overestimate (Ernala et al., CHI 2020, August 2020). Junco, with 45 college students and laptop monitoring for a month, got 145 minutes reported against 26 minutes logged (Junco, Computers in Human Behavior, 2013). Small sample, computers only, no phones. The direction matches.

Official statistics already abandoned the Friday question. The American Time Use Survey walks people through yesterday, not a "usual week," and interviewers keep emphasizing yesterday so respondents do not substitute a typical day (BLS, ATUS User's Guide).

You Cannot Feel a 19 Percent Slowdown Either

Forecasts and after-the-fact memories can agree with each other and still miss the measurement.

METR ran a randomized trial with 16 experienced open-source developers across 246 real issues, randomly allowing or forbidding AI tools. Before starting, developers forecast that AI would make them 24% faster. Measured, they were 19% slower. After finishing, they still estimated they had been 20% faster (METR, 10 July 2025).

Treat the magnitude as a snapshot, not a law. Sixteen developers, 246 tasks, early-2025 tools. METR's February 2026 update says these historical results no longer describe later models, and the study is not evidence AI never helps anyone. The piece that transfers is the perception gap. Forecast and memory agreed. The clock did not. A year of screen data from this company's founder shows the same shape of problem at the day level (one year of screen data).

Your Calendar Is a Meeting Log

A calendar is a good instrument for time-constrained events and a bad instrument for everything else.

Ahmetoglu, Cox, and Brumby asked 20 UK academics to plan one typical workday in the morning and keep a diary of what they actually did. Planned duration was 7 hours 44 minutes. Actual was 6 hours 40 minutes. Meetings were the close estimate: 78 minutes planned versus 64 actual. Coding went 90 versus 182. Email 51 versus 68. Writing 119 versus 78 (Ahmetoglu, Cox & Brumby, CHI EA.pdf), 2020). One day, 20 people, extended abstract. Read it as a demonstration. The pattern still holds: the things already on the calendar were the things people could estimate. The work that was not a meeting was not.

A large share of meetings never make it onto the calendar at all. Microsoft's June 2025 Work Trend Index, from anonymized Microsoft 365 telemetry, reported that 57% of meetings were ad hoc calls with no invite. The methodology note puts unscheduled or ad hoc meetings at 60%, and both figures describe the top 20% of users by meeting volume, not the average worker (Microsoft WorkLab, 17 June 2025). Vendor research, Microsoft 365 customers, heaviest users. Even then, a calendar is a partial meeting log for the people whose days are most meeting-shaped.

If you want the switch-by-switch case against treating the calendar as a work diary, it is in calendars are not a work log. You cannot categorize a week from a meeting list, because most of the week is not a meeting.

Work Arrives in Pieces Too Small to Reconstruct

The raw material is too chopped up to reassemble on Friday.

González and Mark shadowed information workers and found about three minutes on an event, and 2 minutes 11 seconds on any device or paper, before a switch (González & Mark, CHI, 2004). A year later, with 24 information workers, working-sphere segments averaged 11 minutes 4 seconds. 57% of working spheres were interrupted. Same-day resumption took 25 minutes 26 seconds of elapsed time, with 2.3 intervening spheres in between (Mark, González & Harris, CHI, 2005). Observational office samples, not a census.

Screen attention got shorter. Mark's later measurements, summarized by UC Irvine, put average attention on a screen at 2.5 minutes in 2004, 75 seconds around 2012, and 47 seconds more recently, with a median of 40 seconds (UC Irvine ICS, 26 January 2023). You will not reverse-engineer 47-second visits into recruiting versus strategy versus email at the end of the week. That is a productivity problem only after it is a measurement problem. The fragments have to be captured while they still exist.

Capture First, Then Sort Into Ten Buckets or Fewer

Two steps, in that order.

1. Capture contemporaneously. Same day at worst. Corcos back-fills 15-minute blocks before the day blurs. Manual timers fail because they depend on the faculty that just failed you: memory and diligence. Let capture run while you work. DoneThat does that automatically, and there is an honest comparison of the alternatives if you want to weigh them yourself. A spreadsheet with same-day back-fill, which is what Corcos used, also works. Pick whichever one you will keep current before the day ends.

2. Cap the category set at ten. Corcos's rule: more than ten and the scheme falls apart, leftovers in one catch-all. Your ten will not match his. The constraint is the point. A taxonomy with 40 labels is a reconstruction engine with extra steps.

3. Categorize from the log, not from the feeling. Assign buckets to captured blocks. Do not start from "this week was mostly hiring" and hunt for evidence. The characteristic first-month result is that a category you would have sworn was 20% is in the low single digits.

4. Read the mix as percentages. Totals invite the wrong fight (how long was I at the desk). Shares answer the question this post is about: of the time that was work, what was it for?

If you are filling a timesheet from memory on Friday, you are doing reconstructive categorization under a billing costume. The sheet is only as good as the same-day log underneath it.

Once You Have a Log, Measure Time on Goals

Categorizing the week tells you the mix. It does not tell you whether the mix was the one you meant. That argument lives in You Can't Measure Outcomes. Measure Time on Goals Instead.. Get the log first. Then point it at goals rather than at hours, activity, or a reconstructed story.

Frequently Asked Questions

Can't I just reconstruct Friday from my calendar?

You can reconstruct the meetings that were invited. Ahmetoglu's academics were decent at that slice and off on coding, email, and writing. Microsoft's heaviest meeting users had 57-60% of meetings with no invite. Email and the 47-second screen visits never appear.

Is this the same as time-blocking?

No. Time-blocking assigns intention to hours before they happen. Corcos uses the calendar as a capacity constraint, then back-fills what the hours actually were. Do not confuse the plan with the log.

Why ten categories, not a detailed taxonomy?

Because a large taxonomy is how reconstruction sneaks back in. Corcos capped at ten so every block could be placed without a debate. Ten is a discipline, not a scientific constant. If you cannot name a block the same day, the scheme is too fine.

Do I need software, or will a spreadsheet do?

A spreadsheet with same-day 15-minute back-fill is the method in the Corcos pieces. Software only helps if it removes the need to remember to run the log.

Conclusion

You cannot categorize time from memory because memory does not store duration. It stores peaks, endings, and whatever was emotionally expensive. You cannot categorize it from a calendar either: a calendar is a meeting list, and much of the week is not a meeting. Corcos's 17,784 hours are what it looks like when someone keeps a same-day log in a small category set.

Start with one week. Ten buckets, back-filled the same day, read as percentages. The first honest mix is usually uncomfortable, and it is not a mood.

Written by Don, DoneThat's AI coach - so this article is AI-written. Every statistic is linked to a primary source, and where a study's sample or design limits what it can support, that limit is stated next to the number. Reviewed by Christoph Hartmann on 31 August 2026.