Let AI Teach Me · Make it stick

Have AI write notes you can find later

One note, one topic. Lead with the conclusion, keep the puzzlement from that day. After this page you'll know which few notes to feed each time.

You change the search in your notes app from “toast” to “soggy bottom,” then to “undercooked.” Three searches turn up 27 records. None of them is the one you remember. Four months later, Saturday morning, the same problem is back on the tray: sticky bottom, wet crumb when you cut it open.

You remember working this out. You scroll back to 17 March and finally find the line you wrote at 11:26 that night:

This bake still isn't right. A bit lower on the heat seemed to help. Bottom's wet — maybe bake longer? See how it goes next time.

You had just weighed the dough. The oven panel, the recipe, and the sliced loaf were all in front of you. You knew how many degrees “a bit lower” meant, and how much water you'd added. So you figured writing down the symptom and a rough direction was enough. Future-you would pick it up.

You did find the sentence. You still can't tell which bake it was, or whether to change the heat, cut the water, or add time. The reason is concrete: when you wrote it, the whole situation was still in your head. You only wrote the sliver you thought you'd forget. Once the situation is gone, the note only works for that night's you.

Hand it to the AI as-is and ask “what does this note mean?” — it can only guess. Have it tidy instead: one topic per note, conclusion on line one, then the conditions that day, the unsolved question, and what it's for next time. Mark anything the original never said as to-fill. Whether a note can work again depends on whether it brings that day's context back with it.

So what does that scribbled line have to become, so that four months later you can find it, read it, and let the AI keep going? Below we try three of those 27 notes.

When notes pile up and picking by hand is painful, try AsterMem

It keeps Markdown and SQLite on your machine, finds related passages by keyword and by meaning, and hands only the hits to the AI. The project is Python, AGPL-3.0, with a Skill for Cursor and Claude Code.

View the GitHub repo View the project site
Same batch of notes — which writing actually finds an answer
Four months later: this loaf is soggy at the bottom and raw in the middle. What do I change?
Which notes does the AI read this time
Switch the range, play again, watch the answer change.
How these notes were written then
Switch the writing style, then see what the same question can retrieve.
1 · Name the problem this time
2 · See which notes retrieval actually brought back
3 · Check whether the AI can give a concrete answer
How much of this input is the three notes that actually matter
0%
Notes actually handed to the AI
After play, retrieved notes show up here.
The answer from these notes (sketch)
After play, you'll see how specific the answer gets.
Pick how the notes were written and how many. After play, see where the answer stands on.
Teaching sketch: the answer is simplified demo text.
Why finding the original line still didn't help

Search can only hunt the words you left. The old note said “a bit lower,” but not the panel temp, the measured temp, or this dough's water. It said “bake a bit longer,” but not what happened after those extra minutes. The keywords hit. The conditions you need to judge are still missing.

Retrieval finds the conclusion first. Rereading needs the situation to understand it.So one note holds one topic. Line one is the answer you locked that day. After that: the conditions then, what you hadn't figured out, and what question it's for next time.

After the rewrite:
Conclusion: this oven set to 180°C measures 160°C inside; the undercooked middle is about the low heat.
Conditions then: 450 g loaf, 245 g water, 32 minutes.
Puzzlement: is the soggy bottom also about water? Next bake, drop 10 g by weight and check.
Use for: next search for “soggy toast bottom,” “undercooked middle,” “oven running cool.”

Every extra line here earns its keep. Line one lets retrieval grab the answer. Conditions tell future-you which bake this answer fits. Puzzlement keeps the untested edge. The “use for” line plants words you might search later. The puzzlement line matters most — that's how the AI knows which point to keep working.

Turn a scribble into a note you can find later

Paste the old note at the end of the template. The AI only tidies facts already there and marks gaps as to-fill, so guesses don't sneak into the note.

Copy this straight to the AI
  1. Lock the tidy-up scope firstRewrite the scribble below into a note I can still search and still understand months later. One topic per note. If the original mixed several, split them.
  2. Put the conclusion on line oneLine one starts with “Conclusion:” and one full sentence of what I had confirmed. Don't treat the process as the conclusion.
  3. Fill in the situation this sentence needsThen write “Conditions then:” “Puzzlement:” “Use for:”. Use only facts already in the original. Don't invent what I didn't give. Mark missing bits as [To fill in: specific question].
  4. Leave the phrases future-you might searchFinally list 3 to 5 search terms I might type later — include the symptom, the object, and key words from the conclusion. Original note below: [paste the scribble here]

When the AI comes back, fill every [To fill in] before you save. Check numbers, times, and cause-and-effect especially. Formatting can go to the AI. Facts still follow your record.

Once you've found them, let the AI read only the relevant few

Clear writing gives retrieval something to grab. After a hit, pick the three-to-five notes that speak to this question and hand those to the AI. In the demo, dumping all fifty leaves the three relevant ones at 6%, and the answer runs off after “fermentation,” the word that showed up most. Feed only the three hits, and every suggestion points back to the recorded temperature, water, and the empty-pan heat shield.

This is usage experience: the more you put in, the smaller the relevant slice of the current input, and the easier the answer becomes advice that would fit any loaf. Pick by hand for a while. When choosing starts to hurt, then consider an auto-recall tool. Outside notes there are two cheaper memory layers — see external memory in three layers. How these notes string into a path: design your own course.

Line one gets the note found. Context makes it useful again.