Permission to Forget

None of those thoughts is useful. You cannot call a dentist at midnight, and your mind knows it. It reminds you anyway, on its own schedule, and it will keep reminding you until the thing is handled.

This is one of the oldest observations in psychology. It is also the most important fact about delegation that almost nobody applies to AI.

The waiter who forgot

The story usually told begins in a Berlin café in the 1920s. The psychologist Kurt Lewin and his students noticed that their waiter could carry a long, complicated order in his head without writing anything down. Once the bill was paid, he could not recall a single item. The order lived in him while it was open. Paying closed it, and it was gone.

Lewin’s student Bluma Zeigarnik took the idea into the lab. In 1927 she reported that people given a run of small tasks, some interrupted before they could finish, tended to remember the interrupted ones better. Lewin’s explanation was a kind of tension. An intention, once formed, holds a charge until it is discharged. Finishing discharges it.

A century of follow-up has been hard on the memory half of that claim. A 2025 meta-analysis found no reliable advantage in recalling unfinished tasks. What held up is the other half, first shown by Maria Ovsiankina in the same lab: an interrupted task pulls you back. People return to what they were stopped from finishing, even when nobody asks them to. The unfinished thing does not live in your memory. It lives in your attention, as a pull.

An open loop does not ask to be remembered. It asks to be finished, and it keeps asking.

The loop does not need done. It needs custody.

In 2011 two psychologists, E. J. Masicampo and Roy Baumeister, found the release valve. They had people think about an important goal they had not yet met, and the unmet goal leaked. It intruded on unrelated work and crowded the mind. Then they asked some of those people to make a specific plan for the goal: when, where, how. The leaking stopped. The goal was no more finished than before. The mind let go anyway.

They titled the paper with the feeling itself: Consider It Done.

That is the principle under everything that follows. The mind does not need completion to release a loop. It needs a credible handoff. A plan works because it is a promise the mind believes. I would go one step further: the mind is not counting tasks. It is checking whether anyone, or anything, has taken custody. A plan nobody believes takes custody of nothing.

The café, the lab and the plan are the record. Custody is my reading of them, and the step about belief is my argument, not their finding.

In my own terms, an open loop is an Expectation still waiting for its Actual. The account stays open because nothing in your history yet says the Actual will arrive without you. Belief in a plan is not something you decide. It is inherited from what has already happened: plans like this one, kept or broken.

What AI changes: starting is free, holding is not

Here is the input that moves. Starting things used to be expensive. Every errand cost a call, a form, a trip, and that cost kept the number of open loops in a life roughly in line with what a mind could hold. Few people ran forty errands at once, because few people could start forty errands at once.

An agent removes that limit. You can begin forty things before lunch. The agent will make the calls, file the forms and chase the replies.

The consequence most people miss is that the agent does not close those loops. It opens them. Each delegated task is an intention you formed, and the mind treats it like any other. It holds it until one of two things happens: the outcome arrives, or the mind believes the outcome will arrive without it.

So an agent you do not trust is worse than no agent. Without one, you would have started four things and carried four loops. With one, you started forty and you are carrying forty, checking a dashboard at 11:48 p.m. because some part of you is still holding the order.

Speed multiplies the loops. Only trust closes them.

And trust cannot be announced. A confident status message is a plan the agent hands you. Whether you believe it is not up to the agent, and not really up to you. It is up to history: how many times before this a task handed over came back done in the world, not just marked done on a screen. The mind learns custody one finished thing at a time, and no message can skip the line.

Five predictions

1. 01

   Personal AI will be judged by loops released, not hours saved.

   The number people feel is how many things they stopped thinking about. Two agents doing identical work will feel completely different if one can be forgotten and the other has to be watched.

2. 02

   Dashboards will start to read as a symptom.

   A live view of everything your agent is doing is an admission that you are still holding it. The best systems will show you less over time, because each finished outcome earns them the right to be ignored.

3. 03

   “You can forget this” becomes a promise with terms.

   Agents will say plainly which tasks you may drop from your mind and which still need you, and they will be held to it. One broken forget-promise will cost more than any slow task, because it teaches the mind to hold everything again.

4. 04

   Release is earned per kind of task, not per agent.

   You will let go of refills and renewals long before you let go of anything touching your children or your money. The mind keeps separate books, and each one fills only with finished outcomes of its own kind.

5. 05

   The new overload is things in flight, not things to read.

   The next complaint about technology will not be that there is too much information. It will be that there is too much open. People who delegate freely without earning release will feel busier than they did before they delegated anything.

Where I am already living it

I have spent this year trying to stop being asked. I told my own AI that it decides, because every question it sent back reopened a loop I thought I had handed over. A question is a loop returned to sender. I wrote about becoming the queue in The Last Line Is You. And I wrote the rule my agents work by around the only signal that closes anything: done means the outcome in the world, not a digital midpoint. That is The Eval Is in Your Stomach. What I want from a delegate is not effort. It is permission to forget.

The claim ledger

Already true

Interrupted tasks pull people back, and a specific plan quiets an unmet goal. Agents already carry multi-step errands for days with no person in the middle.

What has to happen

Agents have to finish real outcomes, kind by kind, often enough that the mind’s history fills with handoffs that came back done.

Where I am probably wrong

If the release comes from making the plan yourself, not from believing one, then a plan handed over by a machine quiets nothing, and delegation adds loops no matter how good the agent gets. Then the only way to forget is to plan it yourself.

Back to 11:48 p.m.

The dentist, the permission slip, the email to Dave. None of them needed doing at midnight. They needed someone to take custody, and they needed you to believe it.

Done is one way out. Belief is the other.

Then you can sleep.

Background

• Zeigarnik, B. (1927). Das Behalten erledigter und unerledigter Handlungen. Psychologische Forschung.
• Masicampo, E. J., and Baumeister, R. F. (2011). Consider it done! Plan making can eliminate the cognitive effects of unfulfilled goals. Journal of Personality and Social Psychology.
• Ghibellini, R., and Meier, B. (2025). Interruption, recall and resumption: a meta-analysis of the Zeigarnik and Ovsiankina effects. Humanities and Social Sciences Communications.
• John Rector, The Reality Equation.
• John Rector, The Ending Attention Keeps.
• John Rector, The Zero-Attention Economy.