Anki as an Association Engine

Wilhelm Treffer — Fri, 11 Sep 2026 00:00:00 GMT

I could memorise the definition of overfitting and still overfit to it.

In machine learning, a model overfits when it learns its training examples so closely that it performs well on those examples but poorly on new ones. I notice something similar in my flashcards: I can answer a card instantly, then struggle with the same idea in a different setting. I know what comes next on the card. Outside it, I’m less sure.

For ten years, I’ve used a flashcard app called Anki to keep more of what I learn.¹ What interests me is the gap between remembering an answer and being able to use the idea. To bridge that gap, I can change what the question asks me to recognise. A card asking “What is overfitting?” could become:

I can recite a definition when I see my Anki card, but I miss the same idea in a new situation. What machine-learning concept does this resemble?

Overfitting. Now the concept describes something I actually do. To write this question, I have to find an example in my own life. When I review it, I practise recognising the idea in that example. The next time I can answer a card but struggle to use it, I have a possible explanation—and a reason to change the question.

The connection works both ways. When someone describes a model learning its training examples too closely, I have something familiar to draw on: “Oh, that’s what happens with my cards sometimes.” My experience helps me understand the concept. The concept helps me improve the very habit I’m using to learn it.

This is why I think of Anki as an association engine. The questions I choose help shape what will bring an idea to mind. A better question connects the idea to something I might encounter in my life. I’m programming the probability: if this comes up, I’m more likely to think of that. The connection is never guaranteed, but I can give it more chances to happen.

Anki gives these questions repeated chances to do their work. Its loop combines active recall and spaced repetition: trying to remember something, then trying again after time has passed. Both have strong support in learning research. A little effort, spread over years, can keep an idea within reach long after I would otherwise have forgotten it.

That gives me a say in what stays. My work, habits and surroundings already shape what I remember. A card is my vote: this matters; come back to it. As Michael Nielsen puts it:

Anki makes memory a choice, rather than a haphazard event, to be left to chance.

— Michael Nielsen

Some connections come from the cards meeting one another. I keep most cards in one main deck because I want the ideas to mix. A work question from last month might follow a journal note from two years ago, then a passage from a book I’m reading now. Things I learnt separately return together. Each new idea gives them another possible connection. I think of this as a kind of network effect: an idea can become more useful as other ideas give me new ways to understand or apply it.

Meanwhile, my life keeps changing. An idea I’ve reviewed twenty times comes back while I’m facing a new problem. Suddenly, I see a use for it. The card hasn’t changed. What I bring to it has.

Sometimes I notice the connection during a review. Sometimes it comes later. Someone describes a problem, and an idea from a book I read years ago arrives, unasked and exactly right. The thought feels spontaneous. Some of the work happened upstream. By choosing what I learn and return to, I can influence which ideas come to mind later. Keeping those ideas available gives me more material for creative thinking. As Steve Jobs put it:

Creativity is just connecting things.

— Steve Jobs

Once I started noticing how differently an old card could land, I began to see another use for the habit. A card could keep a question alive even when there was no fixed answer to remember. It could become a conversation with my future self, repeatedly drawing my attention to something I want to understand or change.

One of my cards asks:

If I had 10× the agency I have now, what would I do?

Think of a challenging life situation you’re in right now. What would you do if you had ten times the power?

My starting point is usually something like, “I’m doing what I can.” Then I try the question. “Well, if I had ten times more agency, I’d…” And sometimes the next thought is: “Wait, why am I not doing it?” It’s annoyingly effective.

I want to be someone who takes initiative. This question gives me a way to practise: look at the situation I’m in, notice where I’ve been waiting, and see what I could do. Somewhere, I’ve treated the situation as fixed when I could change it. Returning to questions like this may help me develop or change habits. I practise noticing a familiar situation and considering a different response. I have a lot of cards like this. I want them to trigger new thinking. The question stays the same, but my answer changes because I’m changing.

Short journal notes let me revisit that change directly. In a notebook, I rarely return to an ordinary entry. In Anki, it comes back to me: “Oh, that’s what I thought then. I was so young and foolish. Or wait, I was actually on to something.” A few lines bring back an earlier version of myself. The habit gives those notes a life beyond the day I wrote them.

AI writes some cards too. Much of my freshest thinking happens in chats, so a scheduled assistant I call Anki Sensei draws on them to write five cards a day in a separate deck. AnkiConnect, an add-on with a local API, lets a connected AI assistant create or edit cards directly in Anki. I treat the Sensei cards as disposable and move favourites into my main deck.

I still get more from writing cards myself. Choosing the question and shaping the answer are part of understanding. When a card trips me up, I look at the question as well as my memory. Is it vague? Does it ask too much? Do I understand the idea, or have I just become familiar with the sentence?

Choosing also means letting cards go. I archive and suspend them so they leave the queue but remain easy to find. Sometimes an old card shows me how much I’ve learnt. Sometimes it belongs to a question I no longer need to keep asking. The habit has to stay useful and enjoyable enough to last. What I keep returning to shapes my point of view. Over time, I’m writing my own curriculum.

I keep most cards small: one idea at a time. If a card asks three things and I remember two, one rating hides the problem. I split it. If I can repeat an answer but struggle to use it, I try another example. As my understanding changes, I change the cards. Thousands of these small experiments teach me what my mind grabs onto. Each idea I keep can help me understand the next one. Each better question helps me learn it. The knowledge compounds, and so does the skill of working with it.

Writing better questions teaches me about both the subject and how my mind learns. Some skills pay off once. Others change how you learn everything after them. Knowing how your own mind learns is the second kind. No habit has taught me as much about learning as Anki.

Footnotes

The ideas here apply to the spaced repetition system of your choice. ↩

Wilhelm Treffer — Writing

Anki as an Association Engine

Footnotes