Context reconstruction is the work of rebuilding the state of a topic — what was decided, what is still open, and why — before you are able to contribute to it again.
The five minutes before the meeting
You have a call in ten minutes. You know you have discussed this topic before. You know a decision was made. You are fairly sure something was left unresolved.
So you open the transcript from the last session and skim it.
Then the document, because the detail you want is probably there.
Then Slack, because you remember someone raising an objection.
Then your calendar, to work out when this actually happened.
And about five minutes later: “right — that's where we left off.”
Nothing went wrong here. Nothing was lost, nothing was deleted, no tool failed. Every piece of information was exactly where it should have been.
And you still had to spend five minutes rebuilding a picture you already had once, because the picture itself was never stored — only the raw material it was made from.
That is context reconstruction. It feels like preparation, which is why it is almost never counted as a cost. It happens a few times a day, in amounts too small to complain about, distributed across enough people that it never appears in anyone's numbers.
What the research actually shows
The cost of resuming interrupted work is one of the better-studied areas of human–computer interaction, and the findings are consistently worse than people's intuition.
The 2005 field study by Gloria Mark and colleagues at UC Irvine followed knowledge workers through their actual working days and found that fragmented work is the normal condition rather than the exception. Workers switched tasks roughly every three minutes, and returning fully to an interrupted task took an average of 23 minutes and 15 seconds.[1]
Their follow-up experiment in 2008 found something less intuitive and more important. People who were interrupted did not take longer to finish their work — they finished at a similar speed. They compensated by working faster, and paid for it in significantly higher stress, higher mental workload, and more errors.[2]
This is the finding that matters most. The cost of broken continuity does not show up as time. It shows up as strain and as errors — which is precisely why measuring it by the clock understates it, and why teams conclude they are “managing fine” when they are simply absorbing the cost somewhere less visible.
Microsoft's 2023 Work Trend Index, surveying 31,000 people across 31 countries, gives the current scale: the average worker switches between apps and windows more than 1,200 times per day, 68% say they lack enough uninterrupted focus time, and 57% report feeling overloaded by information and communication.[3]
“That research is twenty years old.” Is it still true?
A fair challenge, and worth answering precisely rather than defensively. The honest answer has two halves.
What recent work confirms
The mechanism has held up in controlled experiments published in the last three years.
Hirsch and colleagues, in Memory & Cognition (2023), tested why resumption is costly at all. Across three experiments they found substantial resumption costs and traced them to a specific cause: after an interruption, the suspended goal has to be actively reactivated, and that reactivation takes time. It is not that the task decayed while you were away — it is that returning is itself cognitive work.[5]
Zhang, Kawashima and Shinohara, in Applied Cognitive Psychology (2023), found that recovery speed depends on working memory capacity: people with higher capacity returned to their pre-interruption level of performance faster.[6] Reconstruction cost is not evenly distributed — it falls harder on people who are already loaded.
And Jin and colleagues, in Ergonomics (2024, in print 2025), used multimodal measurement to show that mental workload is highest when the interrupting task requires a different kind of processing from the primary one — and modelled the resulting recovery delay.[7] That is close to the everyday case: leaving analytical work to answer a relational question, then coming back.
The short version: the technology changed. The cognitive cost of switching did not. If anything, exposure went up — the 1,200 daily app switches in Microsoft's 2023 survey are not a number that existed in 2005.
What has not been re-measured — and we will not pretend otherwise
Nobody has repeated Mark's field study in a modern workplace and produced a new headline number. The 23 minutes should be read as the best available field estimate from 2005, not as a current measurement. The recent work above is laboratory work: it establishes the mechanism and its direction rigorously, and it measures resumption in milliseconds on controlled tasks, not in minutes on real projects.
There is a second gap. All of this research studies interruption at the scale of minutes — being pulled away and returning the same morning. Reconstruction across weeks or months, the case that hurts most in advisory, consulting and project work, is the same mechanism at a longer timescale, but has not been quantified in the same way. We think it is very likely worse, because after three months the residual memory that makes a same-day return possible is simply gone. That is a reasoned expectation, not a finding.
So: treat the evidence as establishing that resumption is real, costly, and structural — and treat the calculator below as your own estimate rather than a measurement.
Why it isn't a search problem
The intuitive fix is better search. If reconstruction means hunting through Slack, email and documents, then faster retrieval should solve it.
It does not, and the reason is worth being precise about. Search returns documents. Reconstruction requires a state. Those are different objects, and no amount of retrieval speed converts one into the other.
When you skim four sources before a call, you are not looking for a document. You are performing a small synthesis: reading fragments, discarding what has been superseded, inferring what was decided, and noticing what was left open. The output of that work is a mental model. You then throw it away, and rebuild it from scratch next time.
| You need to know | What search gives you |
|---|---|
| Where does this stand today? | Twelve documents mentioning it, newest first |
| Which decision is in effect? | Both the March decision and the June one that reversed it, equally weighted |
| What is still unresolved? | Nothing — open questions are not a searchable type |
| Why was this chosen? | The conclusion, rarely the reasoning |
IDC's much-cited work on information retrieval put knowledge workers at 15–35% of their time spent searching, with only about half of searches succeeding on the first attempt.[4] That study is from 2004 and its dollar figures have been recycled far past their shelf life, so we cite it only for the shape of the problem rather than as a current measurement. The shape has not changed: retrieval improved enormously in twenty years, and the reconstruction problem did not go away, because it was never a retrieval problem.
Estimate your own cost
Rather than accept an industry average, it is more useful to put your own numbers in. Think about how often in a week you pick up a topic you had put down — a client, a project, a case, a thread — and have to rebuild where it stood before you can contribute.
How much context are you rebuilding?
Two inputs, your own estimates. Move the sliders.
A diagnostic estimate from your own inputs, not a measurement. It assumes 46 working weeks and 7.5-hour days, and it deliberately counts only the rebuilding — not the errors, the repeated conversations, or the decisions made without full context.
Founder, mentor, or running a program? The calculator above counts minutes. In a venture ecosystem the harder question is how many separate, evolving situations you are carrying — and how much of that history would survive if you stepped away.
Take the Context Load Audit → Five questions, three paths, nothing stored.
Most people land somewhere between 40 and 120 hours a year. The number itself matters less than the realisation that it is a recurring line item nobody has ever seen written down.
The four sources of reconstruction work
Reconstruction cost is not one thing. It comes from four distinct failures, and they need different fixes.
1. The topic is scattered across sources
One subject lives in a recording, two documents, a chat thread and a decision someone made verbally. No single artefact holds it, so rebuilding means visiting all of them and assembling the result yourself.
2. Decisions have no status
A decision written as a sentence cannot tell you whether it still holds. The March decision and the June reversal sit side by side as equally valid statements, and resolving the contradiction requires a person who was there.
3. Open questions stop being tracked
A question raised and unanswered is recorded once and then, usually, never carried forward. Months later it is indistinguishable from a question that was resolved — or from one that was never asked.
4. Reasoning is discarded once the conclusion exists
Teams record what they decided far more reliably than why. This is the most expensive loss, because the reasoning is what tells you whether a decision still applies when circumstances change. Without it, people re-decide rather than revisit — sometimes differently, without ever noticing they have contradicted themselves.
Four ways to reduce it
These are tool-agnostic. They are worth doing whatever you use.
- Write the state, not just the event. After a meeting, most people record what happened. The higher-value artefact is two lines on where the topic now stands and what remains open. It takes a minute and removes most of the reconstruction next time.
- Give decisions a status. When a decision is reversed, do not delete the old one — mark it superseded and link the replacement. The pair carries information neither one does alone: that the question was contested, and what was rejected.
- Carry open questions forward explicitly. Keep one list per topic that survives between sessions. An open question that is not carried forward is not tracked, no matter how clearly it was written down once.
- Record the reasoning while it is cheap. One sentence on why an option was chosen over the alternative, written on the day, is worth more than an hour of reconstruction later — and it is the only part that cannot be recovered from the record afterwards.
The honest limitation of all four: they depend on discipline, which is exactly what disappears in busy periods — and busy periods are when the most context is generated. That is the argument for having a system that does it rather than a habit that should.
The same problem now shows up inside the AI
The obvious response to all of this is to hand the reconstruction to a model: give it the entire history and ask where things stand.
That works better than search, and it does not fully solve the problem — for a reason that turns out to mirror the human case.
Gupta and colleagues, at EMNLP 2024, studied what happens when the conversational history given to a language model contains a task switch. Across five datasets and fifteen task switches, they found that sensitivity to history often helps — but that many task switches produce significant degradation in performance.[8] The model had the whole history available. Having it was not the same as being oriented by it.
More context is not the same as better context. A system can hold every transcript you own and still lack a usable representation of where a topic stands — for the same reason a folder full of documents does. Search returns documents. A long context window returns more documents. Reconstruction requires a state, and a state has to be maintained by something.
This is why “just put it all in the prompt” is a weaker answer than it sounds, and why the useful unit is not the volume of material retrieved but whether decisions carry status, open questions persist, and superseded positions are marked as superseded. Those are properties of how the material is structured, not of how much of it fits in a window.
The distinction between a context window and persistent memory is set out in the glossary.
Where continuity comes in — and what that actually means in a product
Reducing reconstruction to near zero requires the state to be maintained rather than rebuilt. That property is what we call cognitive continuity, and context reconstruction is what you pay when you do not have it.
That is the concept. The fair next question is what a tool does about it, so here is ours mapped directly onto the four failures above — each one is a thing the product does, not a benefit statement.
| The failure | What My-CoMind does about it |
|---|---|
| The topic is scattered across sources | Record the meeting, paste the note, or drop in the PDF — that is the entire filing step. Memories about the same topic then gather into one workspace on their own. No tags, no folders, no deciding where something goes. |
| Decisions have no status | Decisions are extracted as items rather than left as sentences. When a later decision supersedes an earlier one, the workspace shows which is in effect and links back to the one it replaced, instead of showing you two contradictory statements. |
| Open questions stop being tracked | Open points are carried in the workspace and persist across sessions until they are answered. Answering one in place is what closes it — nothing quietly falls off the list because a month passed. |
| Reasoning is discarded once the conclusion exists | One recording becomes one memory, so the reasoning stays attached to the decision it produced. Asking “why did we choose this?” returns the reasoning with the source memories cited, and Cognitive Drift shows the sequence of positions with real dates. |
Put together, opening a workspace answers the four questions from the diagnostic at the top of this page — where it stands, which decisions hold, what is open, and why — without you re-reading anything. That is the whole product thesis in one sentence.
Where it does not help. It cannot recover what was never captured. If a decision was made in a corridor and nobody recorded it anywhere, no system has it. What changes is the cost of keeping it: capture has to be cheap enough that it survives busy weeks, which is why the filing step is the part we removed rather than the part we optimised.
And the framing matters more than the tool. Run the diagnostic against whatever you use today — if reconstruction is already cheap for you, you do not need us.
References
- Mark, G., Gonzalez, V. M., & Harris, J. (2005). No task left behind? Examining the nature of fragmented work. Proceedings of ACM CHI 2005, 321–330. doi.org/10.1145/1054972.1055017
- Mark, G., Gudith, D., & Klocke, U. (2008). The cost of interrupted work: More speed and stress. Proceedings of ACM CHI 2008, 107–110. doi.org/10.1145/1357054.1357072
- Microsoft. (2023). Work Trend Index 2023: Will AI fix work? Survey of 31,000 respondents across 31 countries. microsoft.com/worklab
- Feldman, S., & Sherman, C. (2004). The high cost of not finding information. IDC White Paper #29127. Cited here for the proportion of time spent searching; its cost estimates are widely recirculated without re-measurement and should be treated with caution.
- Hirsch, P., Moretti, L., Askin, S., & Koch, I. (2023). Examining the cognitive processes underlying resumption costs in task-interruption contexts: Decay or inhibition of suspended task goals? Memory & Cognition, 52(2), 271–284. doi.org/10.3758/s13421-023-01458-8
- Zhang, H., Kawashima, T., & Shinohara, K. (2023). Interventions to reduce the negative consequences of interruptions on task performance and individual differences in working memory capacity. Applied Cognitive Psychology, 37(6), 1328–1340. doi.org/10.1002/acp.4126
- Jin, H., Liu, L., Luo, Z., Meng, S., & Zhao, Y. (2025). The effects of different interruption conditions on mental workload: an experimental study based on multimodal measurements. Ergonomics, 68(8), 1274–1292. First published online September 2024. doi.org/10.1080/00140139.2024.2400129
- Gupta, A., Sheth, I., Raina, V., Gales, M., & Fritz, M. (2024). LLM task interference: An initial study on the impact of task-switch in conversational history. Proceedings of EMNLP 2024, 14633–14652. aclanthology.org/2024.emnlp-main.811
Sources [1]–[4] establish the scale of the problem in real work environments; [5]–[7] are recent controlled studies confirming the underlying mechanism; [8] extends the same question to language models. Where the evidence stops, we have said so in the text rather than extrapolating.