How to use AI to actually remember what you read

Stop Forgetting Books: AI That Makes Reading Stick

Reading more is not the lever; recalling is. Pair a capture tool like Readwise Reader with a spaced-repetition engine such as Anki running the FSRS algorithm, and use Claude to turn highlights into questions you must answer from memory. Retrieval, not rereading, is what survives the forgetting curve.

I have read books I cannot summarise a week later, and for years I blamed my memory rather than my method. The fix is not a better reading app or a faster pace; it is changing what you do after the page. Hermann Ebbinghaus mapped this in 1885: without reinforcement, you lose the bulk of new material within days, a curve that the 2015 Murre and Dros replication confirmed almost exactly. Reading harder does not flatten that curve. Retrieving does.

The most important study to internalise here is Karpicke and Blunt, published in Science in 2011. Students who read a text and then practised recalling it from a blank page outperformed students who reread and built elaborate concept maps, and they won not just on simple recall but on inference questions that demanded real understanding. The students themselves predicted the opposite, expecting the elaborate method to work better. That gap between what feels productive and what actually builds memory is the whole problem, and it is the gap AI can either widen or close. As an architect I think of this as a pipeline; as a coach I notice that the comfortable option and the effective option are rarely the same, which is exactly the lesson I keep returning to in my essay on discovery.

So here is the setup I actually run. Capture lives in Readwise Reader, which at roughly ten dollars a month ingests articles, PDFs, newsletters, and ebooks, holds your highlights in one place, and resurfaces them on a spaced schedule so old ideas reappear before you forget them. Its built-in AI, Ghostreader, will answer questions about a document and draft flashcard prompts from your highlights. That last part matters, because a highlight is not a memory; it is a bookmark for one. The work of turning it into something you own happens when you try to reconstruct the idea without looking.

For material I genuinely need to keep, I push the hard parts into Anki running the FSRS scheduler. FSRS, built by Jarrett Ye and now the default option inside Anki, models your individual forgetting rate and schedules each card for the moment you are about to lose it, which in practice means roughly twenty to thirty percent fewer reviews than the older SM-2 algorithm for the same retention. The discipline is to write cards that demand a real answer, not recognition. This is where I use Claude as a thinking partner rather than a summariser: I paste a chapter and ask it to interrogate me, one question at a time, refusing to show the answer until I have committed to mine. Ethan Mollick, in Co-Intelligence, calls this the cyborg pattern, where you weave the model into the work itself instead of handing the work over. A summary you read is the centaur move and it feels efficient; being quizzed until you stumble is the cyborg move and it is what sticks.

The failure mode I see most in founders and executives is using AI to avoid the effortful step entirely. You feed a book to a model, read a tidy three-paragraph digest, and feel informed. You are not; you have outsourced the one cognitive act that creates memory. The summary lives in the model's context window, not in your head, and next quarter when you need that idea in a board meeting it will not be there. AI summarisation is excellent for triage, for deciding what deserves your attention, and genuinely poor as a substitute for retrieval. Use it to choose what to learn, never to pretend you have learned it.

There is also an honest limit worth naming. Spaced repetition works on discrete, well-formed knowledge: definitions, frameworks, a founder's specific argument, the mechanism behind a result. It does not work on vague vibes, and writing a good question is itself the bottleneck, which is precisely why having a model draft and pressure-test your cards removes the friction that kills most people's systems by week three. The tool lowers the activation energy; it does not replace the act.

If you want one change that compounds, make it this: after anything worth remembering, close the source and write, in your own words and from memory, the three claims you would defend if challenged. Then let Claude poke holes in what you wrote and let Anki schedule the gaps. The reading was never the work. The reconstruction is, and AI is finally good enough to make the reconstruction almost effortless to set up while keeping the effort exactly where memory is built.