The surprise: most tagging makes it worse
A 2026 study pulled apart into three pieces what is usually lumped together: metadata, structure and retrieval strategy. Across six task sets, with four models from three different model families. The study isolated the third factor, retrieval strategy, as well, but this article is about the other two, because those are the ones note-taking can change.
| What was added | Effect on accuracy |
|---|---|
| Structure only (ordered records out of raw text) | −0.008 … −0.078 |
| Temporal metadata (when it holds) | +0.220 on the task most affected |
| Full enrichment (everything together) | +0.196 on one, −0.032 and −0.063 on the others |
The study's summary sentence is short: most enrichment reduces accuracy.
The one element that works cleanly is temporal metadata — and not by a little: on the task with the best result it accounts for more than a hundred percent of the total gain. Which means every other added layer eats back into it.
Adding a reliability score ("source quality: medium"), a contradiction flag and a provenance chain hurt on every single task. The authors even gave the phenomenon a name: the model processes the metadata correctly and still gives a worse answer.
Translated into note-taking, this is the thing to remember: the date demonstrably helps. Tagging of the "reliability: medium", "X said this at a talk", "source: second-hand" variety can measurably hurt. Completeness is not the same as usefulness.
What measurably helps: the chunk that says where it came from
There is another direction that does consistently improve things, and it is worth telling apart from the previous one.
When the system cuts your notes into chunks, a chunk on its own is often meaningless: "this figure grew by 12% compared with the previous quarter" — but which figure, which quarter? If a sentence about where it comes from is put in front of every chunk, retrieval measurably improves.
According to one vendor's measurement, the share of relevant chunks that were not retrieved fell from 5.7% to 3.7% from that single step, and down to 1.9% together with further methods. (Vendor blog post, not peer-reviewed, the test set is not public — accept the direction, treat the exact numbers with reservations.)
The difference between the two cases matters. The date and the "what this section is about" are information needed to understand the text. The reliability score and the provenance chain, by contrast, are a judgement about the text — and the model measurably copes worse with those.
A warning about numbers of the "31% improvement" kind
This one is worth learning, because it applies to any tool demo.
Rich metadata tagging was measured on a financial question-answering system. The result, in four rows:
| Metric | Naive setup | With metadata |
|---|---|---|
| F1 | 32.9 | 43.2 |
| Claim recall | 45.9% | 39.9% |
| Context precision | 20.0% | 18.0% |
| Faithfulness to the source | 76.4% | 73.7% |
The "31% improvement" is true — in the first row. The other three got worse, faithfulness among them.
If a tool's maker shows you one improving metric, ask about the rest. This is not a point about that product, it is a point about every demo of this kind.
How big a chunk, how big a note?
There is no universal answer, but there is a usable rule of thumb, backed by two measurements:
- For short, factual questions (what was the deadline, how much was the amount) the small chunk is better: roughly 64–128 tokens, which is two or three sentences.
- For broad questions looking for connections (why did we decide this way) the larger chunk is better: 512–1024 tokens, which is one or two paragraphs.
Put into practice: the four-to-eight-sentence paragraph is a usable size in both cases, and three thousand words of unbroken prose is usable in neither.
"I will just drop it into the million-token window" — why that does not work
This is the most common objection to structuring notes: why cut anything up if the model will read a million tokens anyway?
The toughest measurement on this tested 13 models, all of them advertising a window larger than 128,000 tokens. The trick in the measurement is that they removed the literal word overlap between the question and the information sought — so real connections had to be recognised, not keywords found.
The result: at 32,000 tokens, 11 of the 13 models did not even reach half of their own short-context performance. One concrete example: from 99.3% down to 69.7%.
Below a thousand tokens every model was close to flawless.
On top of that comes the older, well-known finding: models make worse use of information placed in the middle of the context than of information at the start or the end.
So the most common misunderstanding is this: a million-token window does not mean it is equally accurate across a million tokens. It means it does not stop with an error.
Which one when:
| Retrieval | Long window | |
|---|---|---|
| Better when | a large body of rarely changing material, a narrow question | one document that makes sense as a whole |
| Cost | cheaper | substantially more expensive |
| Risk | pulls up the wrong chunk | measurably degrades above 32K |
The human side: the myth about the pen
The best-known note-taking claim on the internet is that you learn better writing by hand than typing. It comes from a 2014 experiment.
The most thorough replication did not support it. In the authors' own words, performance did not differ consistently between any of the groups, and pooling the direct replications also showed only a small, non-significant difference in favour of handwriting. And in the experiment where participants were allowed to review their notes, the differences shrank further.
The authors write it themselves: deciding which method is better is premature for now.
What has been measured about AI-made notes, on the other hand, is more nuanced, and not what you would expect:
- In a randomised experiment, an AI-written text summary did not produce worse comprehension than a human-written one, and in delayed recall it even showed a more favourable pattern. Human-written summaries, however, gave better self-knowledge about what someone had actually understood.
- The key correlation in a survey of 319 knowledge workers: the greater the trust in the AI, the less checking — and the greater someone's confidence in their own knowledge, the more.
- A small study that also measured brain activity found that people working with AI could not accurately quote their own text they had just written. (54 people, a single task type, preprint — a signal, not proof.)
The synthesis: the data does not say AI-made notes are harder to remember. It says they give worse self-knowledge about what you understood. Which is why the goal is not to have it "write instead of you", but to get the raw material filed away retrievably — and keep the checking with you.
The how-to
1. DATE IN THE TEXT, NOT JUST IN THE FILENAME
This is the one metadata field that measurably
helps. "Meeting of 12 September 2026" — into the
first line of the note.
2. A HEADING OVER EVERY UNIT OF THOUGHT
Chunking uses this as its cut point.
3. PARAGRAPH = FOUR TO EIGHT SENTENCES
Works at both chunk sizes. Unbroken prose works
at neither.
4. LET THE FIRST SENTENCE SAY WHAT IT IS ABOUT
Do not make it emerge halfway down. The chunk
should stand on its own.
5. DO NOT TAG IT WITH JUDGEMENTS
"Reliability: medium", "source: hearsay" — these
measurably hurt. If something is doubtful, write
it out in a sentence, in the text.
6. NEXT TO EVERY NUMBER, SAY WHAT IT IS OF
"Grew by 12%" on its own becomes a useless chunk.
7. CHECKING DOES NOT GO TO THE MACHINE
Whatever you later claim from it, you look at.
The first and the fifth point together give the point of the article: write fewer labels and more dates.
Common misconceptions
- "The more metadata, the better." The measurement shows the exact opposite: most enrichment hurt, and only the temporal data helped.
- "Always hand it to the model in JSON." Format does matter — in one measurement it made a 40% difference on one task — but there is no single best format, and larger models are less sensitive to it.
- "I remember it better writing by hand." The best replication did not support that. It does not mean writing by hand is bad — it means this is not the deciding factor.
- "I can drop everything into the million-token window." At 32,000 tokens 11 of the 13 models tested already failed to reach half of their short-context performance, once there was no literal word overlap.
- "If it improved 31%, it got better." Look at the other metrics. In that same measurement faithfulness and precision both got worse.
What to take away
A good note is the same thing for the machine as it is for you: broken up, dated, and its pieces understandable on their own. Those three things make a measurable difference, and all three are free.
What is worth stopping, on the other hand, is judgement tagging. What is a verdict about the text, leave outside the text — by the measurement, it does not make the model more cautious, only less accurate.
How this turns into a working, searchable knowledge base is what the a personal knowledge base post covers step by step; and the right order for reading and checking is laid out in the reading and researching faster post.