Zero-Effort Capture, High-Effort Recall

In July 1945, three months before the first stored-program computer ran, Vannevar Bush published a description of a machine he called the memex. A desk that would hold a person's books, records and communications, mechanised so that they could be consulted with exceeding speed and flexibility.

Almost every specific prediction in that essay has been delivered. Storage is effectively free. Consultation is instantaneous. The corpus is not a desk but a planet.

The one thing Bush got wrong, or rather assumed away, is the thing that turns out to be the whole problem. He assumed that a person with a well-stocked memex would consult it. He describes a researcher following a trail through his own material with obvious pleasure, and never once asks what makes the researcher turn to the desk in the first place.

Video. Douglas Engelbart and team, 1968. The demo that delivered most of Bush's mechanism, twenty-three years later. Published by SRI.

Fig. 1. Engelbart's prototype mouse. Built to make consultation cheap, which it did. Photograph by Michael Hicks, CC BY 2.0, via Wikimedia Commons.

The best-funded possible version

If you want to know whether a product category works, do not look at the startups. Look at the strongest version anyone will ever build, and ask what happened.

For "capture everything and find it later," that version is Microsoft Recall.

Recall is not an application. It is an operating system feature, announced at Build in May 2024, delayed for most of a year after a security backlash, and released on Copilot+ hardware. It takes a snapshot of the screen every few seconds and makes those snapshots retrievable using on-device models.

The hardware floor tells you how seriously it was meant:

Table 1. Microsoft Recall system requirements. Source: Windows Recall, Wikipedia.

Microsoft did not build this as a feature. It built a silicon requirement for it, and then shipped hardware to meet the requirement.

And the capture side is, by any measure, solved. Better than solved. It is automatic, continuous, and demands no decision from the person at any point in the day.

So why does the pitch still feel unfinished?

I want to be careful here, and mark clearly what follows as argument rather than measurement. I have not been able to independently verify any figure on how often people actually use Recall's retrieval, and I am not going to assert one.

What I can argue is structural, and I think it applies far beyond Microsoft.

Recall removes the cost of capture entirely and leaves the cost of recall exactly where it was. Arguably it raises it, because the corpus is now everything you have ever looked at, including thousands of frames of nothing.

THE INVERSION

  capture_cost  →  0
     no decision, no gesture, no moment,
     no awareness that anything happened

  recall_cost   →  unchanged, or higher
     you must remember a thing existed
     you must remember to ask
     you must describe it well enough to match

  The system remembers everything.
  The human still has to remember to ask

THE INVERSION

  capture_cost  →  0
     no decision, no gesture, no moment,
     no awareness that anything happened

  recall_cost   →  unchanged, or higher
     you must remember a thing existed
     you must remember to ask
     you must describe it well enough to match

  The system remembers everything.
  The human still has to remember to ask

THE INVERSION

  capture_cost  →  0
     no decision, no gesture, no moment,
     no awareness that anything happened

  recall_cost   →  unchanged, or higher
     you must remember a thing existed
     you must remember to ask
     you must describe it well enough to match

  The system remembers everything.
  The human still has to remember to ask

Nothing prompts you at the moment of forgetting. That is the whole problem, and no amount of index quality touches it.

Notice that this is not a search-quality complaint. Grant Recall a perfect index: semantic, multimodal, instant, free. The failure mode is unchanged, because the failure happens before the query. You did not fail to find it. You failed to look.

This is why I think the category's dominant metaphor is wrong. Everyone describes these products as memory. Memory does not work by query. The characteristic experience of human memory is involuntary: a smell, a phrase, a similar situation, and the old thing arrives uninvited and exactly on time. No product in this category has anything resembling that, and the ones that try build notifications, which is a different thing wearing the same coat.

The trap this sets for everyone downstream

Here is where I have to be careful, because the obvious conclusion is the wrong one and I nearly took it.

If the problem is that nothing prompts you, the fix is obviously a prompt. Surface things. Resurface things. Build a trigger. Every product in this space eventually does, and the graveyard is full of them.

I am refusing to build one, for a rule I now apply generally:

Do not invent a solution to a failure you have not experienced inside your own product.

I have not yet had the experience of needing something, having it in my own field, and not thinking to look. I am confident it will happen. When it does, the specific shape of that failure, what I was doing, what would have reached me, what would have been an interruption instead of a help, is worth more than any amount of reasoning I can do about it in advance.

Designing from an imagined failure produces features that solve an imagined problem elegantly. That is most of what ships, and it is why so many products feel like they were built by people who had never needed them.

Three shapes a solution could take, none of them chosen

Since I am declining to build, the least I can do is state the option space honestly rather than pretending it is empty.

Table 2. The fourth row is the one everyone wants. It is also the one that requires the product to watch you work, which is the line this product does not cross.

The third is what I currently ship, and I am clear-eyed that it is the weakest of the four in reach. It has one property the others do not: it never costs anything on the days it does not help.

The open question, left open

I am leaving this unresolved deliberately:

Can a system make something reachable at the moment you need it, without demanding your attention at all the moments you do not?

Notifications say no. They buy reach with interruption. Feeds say no. They buy it with time. Search says no. It requires you to already know.

I do not have an answer. I think most of the value in this category is sitting behind that question, and I think the reason so few people are working on it is that the capture side is vastly more satisfying to build, demo, and fund.

Bush's researcher, following his trail through the memex with pleasure, is the most optimistic figure in the history of information design. Eighty years later we have built him the desk. We still have no idea what makes him sit down at it.

Sources

• Bush, V. (1945). As We May Think. The Atlantic, July 1945. Full text: web.mit.edu/sts.035/www/PDFs/think.pdf

• SRI International. 1968 “Mother of All Demos” by SRI’s Doug Engelbart and Team. youtube.com/watch?v=B6rKUf9DWRI

• Engelbart's prototype mouse, Computer History Museum. Photograph by Michael Hicks, CC BY 2.0, via Wikimedia Commons. commons.wikimedia.org/wiki/File:Douglas_Engelbart%27s_prototype_mouse,_angled_-_Computer_History_Museum.jpg

• Windows Recall. Wikipedia, consulted September 2026, for the announcement timeline and system requirements. en.wikipedia.org/wiki/Windows_Recall

• SRI International. “1968 Mother of All Demos by SRI’s Doug Engelbart and Team.” youtube.com/watch?v=B6rKUf9DWRI

Related reading

• The Graveyard

• What a DJ Set Taught Me About Search

• The Door That Was Never Open

• Never Create a User Action So You Can Measure It

More from the Moc Mallen Studio journal