← Back to blog
memorydatabaseretrievalscreen-free

The secret notebook: when AI looks things up instead of remembering

Romane kept a secret notebook for a week. Then we played at being an AI that looks up its notes instead of relying on memory alone — and discovered that knowing where to look is also a form of intelligence.

Date12 avril 2026
Target age5-8 yrs
Duration45 minutes
MaterialsA blank notebook for Romane, Colored pencils, A week of patience

Monday morning, Romane placed her notebook on the table with the gravity of someone presenting evidence in court.

She'd drawn a lock on the cover. In red. With a key next to it. I wasn't allowed to read it.

What AI doesn't necessarily memorize

There are two ways to be smart. The first: memorize everything. The second: know where to look.

An AI can do both. It has built-in memory — what it learned during training. But it can also consult external documents in real time (files, web pages, notes). In that case, it doesn't "know" — it finds.

The difference matters. What it memorized might be outdated or wrong. What it looks up is more reliable, but depends on what it's given access to.

Romane was going to spend a week building her own personal database.

The activity

The Sunday before, I gave Romane a blank notebook with a single instruction: write down anything that seemed important during the week. No required format — drawings, words, lists. Whatever she wanted.

At the end of first grade she can write, but not fast or for long. Her notebook was mostly drawings labeled with a word or two: a sun with "sunny" underneath, a bowl with "pasta," her PE teacher's name circled with stars.

The following Saturday, we played. The rules: I ask questions about the week. She can only answer by consulting her notebook. Even for things she knows. The notebook overrides memory.

Then, second round: I take the notebook away. Same questions.

What actually happened

With the notebook, Romane pointed to her drawings to answer. The pasta bowl, Tuesday evening. The sun, Wednesday morning. She had everything — not in elaborate words, but in her own pictures.

Without the notebook, answers became fuzzy. "I think it was pasta." "I can't remember the name."

What struck me: she refused to answer without looking, even when she knew. "Those are the rules." She'd understood that reliability came from the notebook, not from memory alone.

I explained that an AI consulting a document does exactly that. It prefers what's written to what it thinks it remembers. And when it's given nothing to look up, it has to rely on its training. Which is less reliable.

Meryl had been following all of this from a distance, busy stacking blocks. When I handed Romane her notebook, he looked up and asked if he could have one too. Not to play the AI game. Because his sister had one.

I gave him a notebook. He drew in it all evening — mostly circles and what might have been a dog. Maybe the same sleeping dog.

To finish

Romane's notebook is on the shelf in her room. She still writes in it sometimes, without me asking.

If an AI consults a document, does it "know" what's in it — or does it just have access to it?

Romane would probably say: knowing where to find something counts as knowing it.

I'm not sure she's wrong.

What the kids said

Romane refused to answer without flipping through her notebook, even for things she knew. 'Those are the rules.' Meryl asked if he could have a notebook too. He drew in it all evening.