A thought shows up at the worst time. You’re in the shower, walking the dog, halfway through someone else’s sentence in a meeting. You tell yourself you’ll remember it. You won’t. Or you open a notes app, and before it takes the thought, it asks you a question: which folder?

That question is where most note systems quietly die. We spent a long time on this problem while building our own notes app, and the research kept pointing at one conclusion: the win comes from capture you trust, not from organization you maintain.

Working memory holds about seven things, and that is the whole problem

In the 1950s, George Miller published a paper whose title psychologists still quote: “The Magical Number Seven, Plus or Minus Two.” Working memory (the part of your mind that holds what you are thinking about right now) has a hard, small limit. Later researchers have argued the practical number is even lower. Either way, it is tiny. Your phone’s clipboard holds more.

Long-term memory is vast, but you don’t control what gets written to it. So every idea, errand, and half-formed plan you try to “just remember” sits in one of those few working slots instead. Hold four of them and there is not much room left to actually think.

Open loops don’t close on their own

Bluma Zeigarnik noticed something in the 1920s, reportedly watching waiters: they recalled unpaid orders in detail, then forgot them the moment the bill was settled. Her studies found the same pattern: unfinished tasks stay active in memory; finished ones get released. This is now called the Zeigarnik effect.

Your brain treats an uncaptured idea as unfinished business. It keeps rehearsing the thing (the errand, the reply you owe, the idea from the shower) because as far as it knows, rehearsal is the only storage available.

Writing it down ends the rehearsal, but only under one condition: you have to believe the note will come back. If some part of you suspects it will vanish into a folder you’ll never open again, the loop stays open and the rehearsal continues. David Allen built the whole Getting Things Done method on this single observation. His version: your head is for having ideas, not for holding them. The system underneath has to earn that trust, or the head refuses to let go.

Filing is a tax charged at the worst possible moment

John Sweller’s cognitive load theory comes from instruction design. Its core claim: learning fails when the presentation of material burns working memory that should have gone to the material itself. He called that wasted spend extraneous load.

A filing decision at capture time is extraneous load in its purest form. Choose a folder. Pick tags. Write a title. Decide whether this is a “project” or an “idea.” Every one of those choices is charged against the same few slots currently holding the fragile new thought. The thought often loses.

So elaborate taxonomies fail twice. At capture, the decision costs the very attention the note was supposed to protect. At retrieval, you must reconstruct a decision your past self made months ago under different assumptions. Was the lease question under Home, Money, or that Moving folder from 2024? You file in one context and search from another. The cabinet was organized; it just wasn’t organized by the person who came back looking.

Retrieval beats organization once search works

The filing cabinet was rational when finding a document meant walking to a drawer. Location was the only index, so you paid the filing tax up front. Search changed that economics. Most of us stopped foldering email years ago; archive plus search won, and nobody misses the folders.

Notes lag behind email here for one reason: keyword search fails on your own past words. You search “car money thing” and the note says “vehicle depreciation.” You wrote it; you still can’t find it, because you don’t remember your own vocabulary from three months ago.

This is the part that has actually changed. On-device language models can now match the question you’d naturally ask against the words you actually wrote, and hand back the note with its source. Once retrieval works at that level, hierarchy stops paying rent. Organization becomes something a machine does after capture, or something you skip entirely.

Three things a trustworthy inbox needs

Strip the research down and the design requirements are short.

Instant capture, from anywhere. Thought to saved note in a few seconds: typed, spoken while walking, or shared straight from another app. Friction is a filter, and it filters out the marginal thoughts first. The marginal ones are often the good ones. (If a blank capture box stalls you, a prompt helps; we keep a free library of journal prompts for exactly that.)

Zero decisions at capture time. No folder, no tags, no title, no “where does this go.” Capture should cost roughly the attention of saying the thought out loud. All organizing happens later, by the system, or turns out to be unnecessary because retrieval is good.

Resurfacing you don’t have to remember to do. This is the one pure inboxes miss. A write-only inbox just relocates the open loop: now “check the inbox” is the thing your brain rehearses. Anything with a date or a deadline must come back on its own, and everything else must be findable by asking in plain words. Trust is built by return, and the Zeigarnik loop only closes when return is reliable.

Notice that none of these are organizational features. That is the shift worth making: judge a note system by how cheap capture is and how reliably things come back, not by the elegance of its folder tree.

The design brief we ended up building

This research is, more or less, the design brief for Second Brain, our notes app. Capture by text, voice, or share sheet, with no folders and no decisions. An on-device model does the organizing after capture instead of you during it. Ask answers questions from your own notes only, cites the note it drew from, and says it doesn’t know rather than inventing. Reminders come back when due. AI processing stays on-device, there is no Studio AM account or processing server, and optional private iCloud note sync is off by default. Devices with Apple Intelligence use Apple’s built-in model; other supported iOS or iPadOS 26 devices may require a one-time local-model download. We wrote about that architecture in A weak model and a strong harness.

But the brief matters more than any app, ours included. Whatever tool you use: make capture instant, defer every decision past the moment of capture, and verify that things actually come back. Your working memory has about seven slots. Spend them on thinking, not on holding.