2026-05-15 · 3 min read
Mnemonics that don't ask you to picture anything
I have aphantasia. When you tell me to picture a red apple, nothing happens. I know what a red apple looks like, I can describe one, I can recognise one — I just can't summon the image. Close my eyes and it's dark.
Around 3-4% of the population is in the same boat, and it's mostly invisible: most aphantasics don't realise other people are actually seeing things when they close their eyes. We assume "picture this" is a figure of speech. It isn't.
It took building a language-learning app for this to matter to me, professionally.
The classical mnemonic problem
Open any vocabulary textbook from the last fifty years and you'll find the same trick: the keyword method, popularised by Atkinson in 1975. Want to remember the Spanish word pato (duck)? Imagine a duck holding a pot on its head. Imagine it vividly. Picture the colours. The image is what binds the sound to the meaning.
It works. For most people. The retention gains over rote are large — Atkinson's experiments showed 2-3× improvements, replicated for half a century.
It does not work at all for aphantasics. There is no image to bind anything to. The instruction is a null operation. We read the suggestion, we understand intellectually what we're supposed to do, and nothing happens. Most language apps quietly fail this slice of their users without ever realising it.
What actually works
The research on aphantasic learners is thinner than it should be, but the directions are clear:
- Decomposition. Chinese is forgiving here: 飞机 isn't a thing you imagine, it's literally fly + machine. The compound itself is the mnemonic, you just have to surface it. Etymology stories do the same job for opaque words.
- Verbal hooks. Puns, rhymes, sound-alikes across languages. Pure linguistic, no imagery. 苹果 (píngguǒ, apple) → "ping-pong fruit." That's a verbal hook, not a picture.
- External visual anchors. This one matters most: an emoji on screen replaces the image an aphantasic brain can't generate internally. The 🍎 next to 苹果 isn't decoration. It is the imagery — externalised so it works for everyone, including the people whose mind's eye is dark.
- Production. Typing the word, saying it, writing it. Embodied memory routes around imagery entirely. We already do this — every review starts as a typed answer.
The technique we ship
For every word in the vocabulary, we generate and cache two pieces, shown inline under the meaning the moment the card is revealed:
- One or two emojis, rendered via Twemoji so 🥟 looks identical whether you're on iOS, Android, or a fifteen-year-old Linux laptop. No more "the dumpling is a square on my system." The emoji is the external visual that does the work an aphantasic mind's eye won't.
- A short verbal phrase — never imagine this, never picture this. Always decomposition, etymology, or a sound-hook. 飞机 — literally a flying machine.
The prompt we use to generate these is explicitly aphantasia-safe. The words imagine, picture, visualise, and in your mind's eye are banned from the output and rejected as failed generations if they slip through. The instruction we give the model is: produce a hook a blind person could use.
A third piece — a longer expansion in a tap-to-open modal, for words where the short hook isn't enough — is reserved in the data model and will surface once we've watched the inline form in the wild for a few weeks. Shipping smaller first.
The rollout
We have thousands of words across multiple languages. Generation is a few seconds per word, but cumulative work runs into days. We process in priority order — words you, specifically, are reviewing get generated first. The rest fills in over the next week or two. If you see a card without a hint, that's why; come back in a few days.
Why this matters beyond aphantasia
Even if you can picture things, an emoji on screen beats an emoji you have to summon. The technique works for everyone — the aphantasia framing is what forced us to do it right. We were ready to ship a generic keyword-method generator. We almost did. The fact that the founder couldn't use his own product is what caught it.
Accessibility constraints make better products. Reliably.
If you've ever read "picture a duck balancing a pot on its head" and felt nothing, you weren't doing it wrong. The instruction was wrong.