AI Writes Notes Faster Than You Can Read Them: Appraise, Don't Hoard

Generating a note is not keeping it. On August 21, 2026,¹ the U.S. National Archives told federal agencies that their "use of AI platforms does not, in and of itself, result in the creation of federal records"² — only later use can. Borrow that test: mark AI output temporary, date a review, keep what you used.

A tool can produce a meeting summary every time a call ends. Reading one is slower. Marc Brysbaert's meta-analysis of 190 studies puts the average adult's silent reading rate in English at 238 words per minute for non-fiction³ — a 1,500-word summary costs about six minutes. Few people spend them. A Hacker News commenter described AI meeting tools in August 2026: "Full transcripts of entire meetings and no human ever takes a 2nd look at them for literally any reason, ever."⁴

The pile is not a failure of discipline. It is a missing decision.

Archivists have a word for that decision — appraisal — and a habit of making it on a schedule rather than at the moment something is created. This essay borrows both. It borrows the reasoning, not the law: the federal rules quoted here govern U.S. agency records, not your notes, and nothing below is legal advice.

Why keeping every AI note feels safe

The case for keeping everything is strong. Storage costs almost nothing, search finds old text in seconds, and nobody knows in advance which note will matter. A 2011 field study found that filing email into folders did not help people find it again. If search works, deleting looks like risk without reward.

Steve Whittaker and four colleagues studied 345 long-term email users who carried out more than 85,000 refinding actions. Their verdict on the time people spent foldering: "these preparatory behaviors are inefficient and do not improve retrieval success."⁵

That finding deserves to be conceded plainly. Sorting AI notes into folders is mostly wasted effort, and nothing here asks you to do it.

The second argument is humility. Catherine C. Marshall of Microsoft Research, writing about personal digital archives in 2008, observed that people lack "the prescience to know which things they will eventually care about."⁶ If you cannot know, keep it all — or so the reasoning goes.

A third argument arrives with the tools themselves. Let the machine read the pile, and you never have to. It is a fair point, and it moves the reading somewhere else. It does not remove the decision — whatever the machine retrieves from is still whatever you kept.

Why the pile still costs something

Keeping has a cost that search does not remove. An unread AI summary sits beside notes you wrote, looks just as authoritative, and may be wrong. Even before AI, a study of office paper archives found only 49% were unique documents. Archivists call the fix appraisal — deciding how long a record stays, by its use.

Steve Whittaker and Julia Hirschberg tracked office workers through an office move. In 2001 they reported that "only 49% of people's archives were unique documents, the remainder being copies of publicly available data and unread information."⁷ The filers among them were also "loath to discard filed information, even when its value is uncertain."⁷ AI output makes that remainder larger and cheaper to produce.

The collector's fallacy explains why saving it feels like progress — the question here is what to do once it exists.

The cost is not disk space. The Hacker News commenter quoted above added that when such a summary finally does get read, it can turn out to say the opposite of what happened in the meeting.⁴ An unchecked summary filed next to your own notes is a quiet source of error.

Archivists met this problem decades ago. Leahy wrote in 1949, as quoted in the Society of American Archivists' Dictionary of Archives Terminology. The line is blunt: "Records of value can now be preserved only by aggressive, planned destruction of the vast majority of duplicated or transitory records."⁸

The dictionary's own definition carries the idea this essay needs. In its second sense, appraisal is "the process of determining the length of time records should be retained, based on legal requirements and on their current and potential usefulness."⁹ The measure is usefulness, current and potential. Not effort spent, not file size, not who wrote it.

Lewellyn, in 1979, as quoted in the same dictionary, named the trade-off without softening it: "The increasing costs of labor, space, and overhead must be weighed with cold brutality against the probabilities of use."¹⁰

What the National Archives decided about AI output

The Archives' 2026 memo ties records to use, not origin. It lists five common AI uses, including summarizing documents and generating meeting summaries, and says output can become a record when the agency relies on it, circulates it, or files it in its systems. Unused output is likely not a record. "The fact that AI was used is secondary."²

Read the list and it describes much of personal AI use. Two of the five entries: "the user asks the model to summarize and/or synthesize information across certain documents" and "the user directs the AI platform to generate a transcript or summary of a meeting."²

The memo then sorts the output by what happened to it. Its list of materials "likely not federal records" includes "AI query results that the agency does not act upon in the course of agency business, such as AI research results that are never used."²

Its first worked scenario follows an employee who asks an AI for early research, changes the focus of her assignment, and never captures or circulates the answers. The memo's conclusion: "The AI materials are likely non-records."²

Origin is not the test. Use is.

The memo also leans on NARA's General Records Schedule 5.2, the standing rule for short-lived records. It says "Transitory records are routine records of short-term value (generally less than 180 days)."¹¹

Intermediary records exist only to produce a later record, and the instruction for them reads: "Destroy upon creation or update of the final record, or when no longer needed for business use, whichever is later."¹¹ An AI summary that fed a note you then wrote yourself fits that second shape exactly.

A boundary belongs here. The memo "addresses records management requirements under the FRA only"² — the U.S. Federal Records Act, as it applies to federal agencies. It does not govern a personal notes folder. If an employer retention policy or a legal hold covers your work notes, that policy wins. This is the reasoning of one profession, borrowed; it is not legal or records-compliance advice.

Decide on a date, not at the moment of creation

Archivists write the keep-or-destroy rule down in advance and apply it later, when use can be seen. For AI notes, that means two moves: label the note as machine-made when it is created, and set the date on which you will decide. The decision itself waits for evidence of use.

The Society of American Archivists defines a retention schedule as "a document that identifies and describes an organization's records, usually at the series level, and provides instructions for the disposition of records throughout their life cycle."¹² The instructions exist before the records need them. Nobody improvises the decision while the box is still being unpacked.

Personal archives work the other way, and Marshall saw where that leads: "because digital stuff accumulates so quickly, it is doubtful that anyone can cull and keep track of all that they have effectively."⁶

Deciding early is no better.

Whittaker and Hirschberg expected careful filers to keep smaller, busier archives. The result: "Contrary to our predictions, filers amassed more information, and accessed it less frequently than pilers."⁷ They called the habit "premature filing" — archiving things to clear the desk that later turn out to be of low value.⁷

The two findings point the same way. You cannot judge an AI summary the minute it appears, and judging it then makes the pile bigger. Ninety days later the evidence is in — you used it, or you did not.

This is not spaced repetition. Tools such as the Spaced Repetition plugin for Obsidian — 630,887 downloads in the plugin registry on November 10, 2026¹³ — bring a note back so you remember it. A review date brings a note back so you can decide whether it stays.

The review line: what to do tomorrow

Three actions carry the rule. Add two lines to the frontmatter of every AI-made note: where it came from and when you will review it. Check those dates inside a review you already do. On the date, keep what you used in your own words and delete the rest — or set a new date.

This is our suggestion, modeled on archival practice. It is not an archival standard for personal notes. In a plain Markdown file it looks like this:

---
title: Q4 planning call (AI summary)
source: ai
created: 2026-11-10
review: 2027-02-08
---

The review date sits 90 days after creation. That horizon is our choice — half of the 180-day line NARA uses as its rough boundary for transitory records. Shorten it if your AI tools run all day.

On the review date, each note lands in one of four places:

1. You used it, and it still holds. Check it, rewrite what matters in your own words, and remove the source and review lines. It is now your note.
2. It fed a note you wrote. It was an intermediary. Your own note is the final record; the summary can go.
3. You never opened it. This is the "AI research results that are never used" case.² Delete it.
4. You cannot tell yet. Keep it and write a new date. Keeping is always an allowed outcome.

A review date is a decision point. It is never an automatic deletion. The difference matters, because a retention rule applied by machine can destroy things people still needed. In June 2026 a Hacker News commenter described a five-year retention rule treated as a five-year deletion rule — and the result: "I have recently wanted to refer back to my notes on some older services and discovered that all my old notes have been deleted."¹⁴

Archivists build the same caution into their own reappraisal practice. Rapport, in 1981, as quoted in the SAA dictionary: "Since what we destroy we cannot retrieve, accessioned records that fail reappraisal deserve safeguards."¹⁵ And Powell, in 1991, quoted in the same entry: "Most archivists are aware that no appraisal decision is perfect."¹⁵

Where the check happens is less important than that it happens. If you already run a weekly review, the due dates belong in its inbox step. A plain text search is enough to find them: grep -rl '^review: 2027-02' notes/ lists every note marked for review in February 2027.

This is the middle path. It is not deleting the whole vault, and it is not keeping everything because saving was easy. Whether AI should write your notes at all is a separate question about memory; this rule starts after the notes exist.

What this rule does not claim

The limits are worth stating first. The federal rules are an analogy, not law for your notes. The research for this essay found no measure of how much AI output people reuse, so the case rests on reasoning and on older studies of paper and email. Search is good — this rule decides what stays, not how you find it.

The memo is also new. It was issued on August 21, 2026, and NARA writes that it "may issue additional guidance as new uses emerge."² Its categories are a tested way of thinking, not a fixed list to copy.

The 90-day date is arbitrary in the honest sense — someone has to pick a number. NARA picked 180 days for federal transitory records. A shorter horizon suits a tool that writes every day.

None of this argues for fewer AI notes. It argues for decided ones.

Frequently Asked Questions

Which AI summaries should I keep? Keep the ones you used. The U.S. National Archives' 2026 memo says AI output can become a record when someone relies on it, circulates it, or files it, and lists "AI research results that are never used" as likely not records. For personal notes, the same test works: use is the evidence that a summary earned its place.

Do I need to save AI meeting summaries? Only the ones that do a job. NARA's memo calls AI meeting summaries likely records when they "are circulated to other employees or that serve as the official meeting record," and likely not records when created "for the user's recollection" and never circulated. If your employer has its own retention policy, follow that policy instead.

Should I delete AI notes I never read? Usually, yes — on a review date, not on impulse. A note you never opened by its review date has shown no use. Deletion cannot be undone, so the review is deliberate and keeping stays an option. As one archivist, Rapport, wrote in 1981, quoted in the SAA dictionary: "Since what we destroy we cannot retrieve, accessioned records that fail reappraisal deserve safeguards."

If search finds everything, why delete anything? Search finds text; it does not check it. A 2011 study of 345 email users found filing did not improve retrieval, so this rule does not ask you to organize anything. The cost of keeping is different: unused, unverified AI summaries sitting beside notes you wrote, retrievable and just as authoritative-looking when they are wrong.

Isn't this just spaced repetition? No. Spaced repetition brings a note back so you remember it; tools built for that, such as Obsidian's Spaced Repetition plugin, are popular for that reason. An appraisal review brings a note back so you can decide whether it stays. One is about memory, the other about what your archive holds.

What should I keep from an AI summary I used to write a note? Your own note. NARA's schedule for intermediary records, which exist only to produce a later record, says to destroy them "upon creation or update of the final record, or when no longer needed for business use, whichever is later." Once your note exists and nothing else depends on the summary, the summary has done its work.

My AI is generating too many notes, how do I manage them? Label them at creation and decide later. Add source: ai and a review: date to each note's frontmatter, check due dates inside a review you already run, and on the date keep what you used and delete the rest. This is a suggestion modeled on archival practice, not a standard.

A machine can write the note. Only use can make it worth keeping.

Notes in MNMNOTE are plain Markdown kept on your own device, so a review: line is just text any search can find — mnmnote.com.