How to Remember What You Read

Here is the test, and you can sit it tomorrow. Read this article the way it tells you to read books: at the end of each section, look away and say back what it claimed. That is closed-book recall, which the research calls retrieval practice, and it is the first of the three parts of the method: recall each section with the book shut, chain or place the chapter's claims so the order survives, and review at widening gaps that start tomorrow and end at one a month. At the bottom are ten questions about the article itself. Do the stopping and you should get most of them tomorrow without looking. Only read and, if the lab numbers carry over, expect about half, and to be surprised, because it will have felt like remembering all the way through.

That feeling is the problem. Finishing a book feels like remembering it; then someone asks what it was about and you produce the title, the mood and one anecdote. Nothing is wrong with your memory. Fluent reading and durable memory are different processes, and the first one lies about the second.

Why you forget a book the week you finish it

The cleanest demonstration is a 2006 experiment by Henry Roediger and Jeffrey Karpicke. Each of 120 students read one short passage twice and read a second passage once, then closed that one and wrote down what they could recall. Every student did both; what differed between them was when the final test came, five minutes later for a third of them, two days for a third, a week for a third. The five-minute group had more of the passage they had reread, 81% against 75%, which is why rereading feels effective. The week group had 42% of the reread passage and 56% of the recalled one, and the recalled passage held at a week roughly what the reread one held after two days (54%): rereading wins the immediate test and loses every later one. And in the second experiment (below), asked to predict their own performance, the students who had only reread were the most confident group.

Karpicke and Janell Blunt ran a harder version in 2011: a science text, recalled from memory or turned into a concept map with the text open, among other conditions. Across eighty students, recall scored 0.67 on a one-week test to concept mapping's 0.45, about a 50% advantage, and it held on inference questions. A second experiment put both conditions inside the same student: of 120, 101 did better after recalling, yet 90 of the 120 had predicted concept mapping would be as good or better. Three quarters believed the weaker strategy was at least as good.

Then there is the speed of the loss. Hermann Ebbinghaus measured it on himself in 1879 and 1880 with nonsense syllables, publishing in 1885, and scored memory as "savings", the share of the original learning effort he did not need when relearning, which is not the share recalled: 58.2% after twenty minutes, 44.2% after an hour, 35.8% after nine hours, 33.7% after a day, 27.8% after two days, 25.4% after six days and 21.1% after a month. Jaap Murre and Joeri Dros published a replication in 2015 with the same shape and lower values (47.2, 37.3, 27.6, 31.7, 23.0, 16.8 and 4.1% at the same intervals), a small rise at the one-day point that the sleep literature would predict, though the authors decline to pin it on sleep, and one large deviation, at a month, where their subject kept far less than Ebbinghaus; they called the replication a success. A real chapter decays more gently than a syllable, but the shape holds: the steep part of the curve is the first day, and that is where the method has to act.

Decide what you want from the book before you open it

In 1972 John Bransford and Marcia Johnson had people recall a deliberately vague passage ("first you arrange things into different groups"). Told beforehand that it was about washing clothes, people understood and remembered it; told afterwards, they did no better than people never told. Context only helps while you take the passage in, so those five minutes are worth more before chapter one than the same five minutes of rereading at the end.

Spend them on two things. Write one question you want the book to answer, with a question mark, something the book could fail to answer. Then survey the skeleton: the contents, the first and last paragraph of each chapter, any summary the author provided. Mortimer Adler and Charles Van Doren called this inspectional reading; Francis Robinson's SQ3R (Survey, Question, Read, Recite, Review) taught it to a generation of students as its first two steps. The question also tells you when to quit. A book that has answered it, or has plainly shown it will not, has earned a blank page and nothing more; finishing it out of duty buys nothing your memory will keep.

The example this article carries to the end is the first chapter of Adam Smith's Wealth of Nations, the pin factory. The question: "What does Smith say actually makes a worker more productive?" The survey: the chapter is called "Of the Division of Labour", opens on the productive powers of labour and the pin manufactory, and closes with the day-labourer's coat and the African king. Now the pin factory is an answer rather than an anecdote.

Read in chunks, then close the book

Robinson's second R is "Recite", and it is the step everyone drops. Read a section, a few pages at most, close the book and say what it claimed in your own words, aloud or on paper. Then open it and check. One minute per section. The stop doubles as the attention check: if you cannot say the section back, you were not reading it, whatever your eyes were doing, so go back to the last paragraph you can. It also sets your pace: read as fast as you can and still pass the stop, and no faster (more on speed reading in the questions at the end). Reading several books at once is fine, and the rule is about where you stop rather than how many books you are holding: switch at a section stop, never mid-argument, so each one is put down in a recalled state rather than a fluent one. What empties a half-read book out of your head is not the company it keeps on the bedside table, it is being abandoned mid-page with nothing retrieved.

Roediger and Karpicke's second experiment gives the dose-response. Students who read a passage once and recalled it three times remembered 61% a week later; read three times and recalled once, 56%; read four times and never recalled, 40%. Five minutes after the session the order was again reversed (71%, 78%, 83%). Every extra reading buys confidence today; every extra recall buys memory next week.

The recall will feel worse than rereading, because it is harder, and hardness is what you are paying for. Do not build images while you read, the beginner's mistake with any memory technique: recall the chunk first, then decide whether anything in it needs an image.

Notes that help you remember and notes that only feel like it

In 2013 John Dunlosky and four colleagues reviewed the evidence for ten common study techniques. Only two were rated high utility: practice testing and distributed practice, which is to say closed-book recall and spacing. Highlighting, rereading and summarisation were rated low, alongside the keyword mnemonic and imagery for text, which is why this article spends images only on the order of the claims and testing on everything else. It inverts what most readers do with a pen.

Highlighting fails because it is a judgement about which sentence matters, made while the sentence is in front of you; it never asks your memory for anything. Rereading fails the same way: Aimee Callender and Mark McDaniel had students read textbook chapters twice in a row, and across four experiments the second reading gave little or no benefit on multiple-choice questions, on summaries, or on the questions that asked students to apply what they had read, for strong and weak readers alike.

Two kinds of note make your memory do something.

The blank page. At the end of each chapter, close the book and write everything you can recall: claims, examples, order. Then reopen it and fill the gaps in another colour. The gaps show where your reading was fluent and your memory was not.

Question-notes. Instead of copying a claim, write the question it answers, with the answer on the back of the card. "Why can farming not divide labour as finely as manufacturing? Because its tasks follow the seasons." A question-note is a test you can sit next week; a copied sentence is a shorter book. For the claims that matter, add three questions and answer those too: why is this true, what does it change, and what does it contradict that you already believed? The why-question is what Dunlosky's review calls elaborative interrogation, rated moderate, below testing and spacing and above anything done with a highlighter; the other two are how a book gets joined to the others you have read instead of being filed beside them.

Both kinds of note outlive the book, so keep them somewhere that is not book-shaped: a file per theme, which is the old commonplace book, or a Zettelkasten if you want the notes to link to one another. Which one you pick matters less than the rule under all of them, in the questions at the end.

Turn each chapter into one image and chain them

A book is an argument in order, and retrieval practice does not reliably give you the order. For that, the oldest tool is the best one. In 1969 Gordon Bower and Michal Clark had students learn twelve lists of ten unrelated nouns, one group weaving each list into a story, the other simply studying. Both recalled each list almost perfectly straight away; on the final test of all 120 words the story group recalled a median 93% and the controls 13%. That is the link method.

Reduce each claim to one concrete image, then join each image to the next so that one pulls the next out. Books are made of abstract nouns, and the fix is Harry Lorayne's substitute-word rule: swap the term for something picturable that sounds like it, then tie the picture to the meaning, the same move as the keyword method for foreign vocabulary.

Smith's chapter, reduced to seven claims after a blank-page recall:

1. Division of labour is the biggest cause of a nation's productive power.
2. The pin factory: one untrained man could scarcely make a pin a day; ten specialists, across about eighteen operations, made around 48,000.
3. Manufacturing divides labour finely; farming cannot, because its tasks follow the seasons.
4. First cause: repetition makes the worker dexterous.
5. Second cause: no time lost switching tasks.
6. Third cause: the focused worker invents machines (the boy who tied a string to the engine valve so he could play).
7. A labourer's woollen coat is the work of thousands, and a frugal peasant is better accommodated than an African king.

Three need substitutes. Division of labour becomes a long loaf cleaved into eighteen slices. Dexterity is a deck of cards ("deck-sterity") shuffled one-handed. Opulence, Smith's word for the wealth that reaches the lowest ranks, is an opal lens.

Now the chain. Out of each slice of the cleaved loaf a pin is pulled, until pins spray across a farmyard where a farmer with a calendar cannot catch them because he must wait for spring. He drops the calendar and shuffles a deck of cards one-handed, and the cards flick into a slot without his turning round. Behind him a boy ties a string from the engine valve to a swing, and the swing hangs from a woollen coat that covers a village, with a peasant on top examining a pin through his opal lens.

Seven images, five minutes. Walk it once now, and again tomorrow before you open the book, writing the seven claims from it. Never made an image before? The ten-grocery challenge is the warm-up: ten items, scored before and after the story, no account needed.

A memory palace for one book: a worked example

A chain holds one chapter. For a book you want for months, a chain of chains will break somewhere, and you need addresses. The method of loci is the address system: known places in a known order, an image at each. Cicero credits it to Simonides in De Oratore, and Joshua Foer's account of memory competitors shows what they put in their palaces: ordered material, decks, digits, lists. A per-book palace should hold the same kind of thing, the claims in order: one room per chapter, one locus per claim, in a building you can walk without effort, which usually means home. Smith's chapter is chapter one, so it takes the first room you enter, the kitchen.

Chapter two goes in the next room, and so on down the hall.

Is it worth it? For most books, no. Mnemonics carry the skeleton and testing carries the understanding: a palace with seven loci should still give you Smith's seven claims in order months later, not his reasoning about why the string matters more than the boy, and if you put every paragraph in a room you will have a house full of furniture and no argument. The image techniques in the technique library all have that shape, holding what can be pictured and put in order and never the argument, and any page offering one of the three parts as the whole answer is offering a third of one. A palace costs ten minutes a chapter the first time and a few minutes per review: right for a textbook you will be examined on or a book you will quote for years, too much for a novel or a one-idea business book, where a question before and a blank page after are enough. Build it when you can name, in advance, the month in which you will need the book without it in your hand.

If you would rather not use your own home, the stock palaces on this site are six illustrated floor plans with their loci already named and placed, or you can draw your own and name each locus as you drop it; either way you fill them with images afterwards, on the palace's own page.

The review calendar: next day, a week, a fortnight, then monthly

The forgetting curve starts running the moment you stop retrieving, however well you encoded. Before you put the book down, walk the chain or the room once and write the claims from it: that is the last recall of the reading session, where you find and repair a weak image, and it is part of building, not a review. The first spaced review is the next day. After that the gap grows: about six days after that first review, then about sixteen days after the second. Those three are SM-2 run forward: one day, then six, then each gap multiplied by the item's ease factor, which clean recalls have raised to about 2.7 by the third review, a little over a fortnight. Left to run, SM-2 would stretch the fourth gap past six weeks; a monthly walk of the chapter is the round number to keep instead, and it errs early. Three spaced recalls and a monthly walk are what a chapter you mean to keep costs. Dominic O'Brien teaches the same widening gap, and names five places to stand in it. In his own words: review immediately and you hold it for a good day; review no more than a day later and you hold it for a week; review a week later and you have it for a month; review a month later and you have it for three to six months; review three to six months later and you should have it for life. Each review buys a longer stretch than the one before, which is the whole idea. The spaced repetition guide compares the named schedules and explains why the gap should grow with how long you want to keep the material.

Rereading has its place here, and only here. Katherine Rawson and Walter Kintsch, in two experiments with 423 students, gave a text once, twice in a row, or twice a week apart. On an immediate test only the back-to-back second reading beat a single reading; two days later only the spaced second reading did, and the back-to-back one no longer differed from reading once. If you read a chapter twice, read it again about a week later, after a recall attempt rather than instead of one.

For the Smith chapter read on 14 September, that is a walk on 14 September before you close the book, then 15 September, about 21 September, about 7 October, and early November. Put the dated ones in your calendar before you close the book.

Two tools on this site do the drilling and the bookkeeping; neither does the reminding. Put the question-notes and key terms into a deck (type them in, or import a CSV) and run it through the custom drill: a timed "memorize" phase shows the cards, a closed-book "recall" phase you grade yourself, then "results", with 5, 10 or 20 cards dealt in SM-2 order: cards whose due date has passed first, then the ones you have never seen, then the rest in deck order. If you built a palace, palace review walks every filled locus in the palaces you pick as flip cards, most overdue first (a locus never reviewed counts as due now), and your easy, hard or miss grade sets its next SM-2 due date. It lists only palaces you have adopted and will not start until the palaces you tick hold three filled loci between them, which the seven-locus kitchen above clears on its own. It does not hide loci that are not yet due and it will not remind you, so the dates above still go in your diary.

Read as if you will have to explain it

In 2014 John Nestojko and colleagues told some readers they would teach a passage to another student and told others to expect a test. Nobody taught anyone; everyone was tested; and the group expecting to teach produced more complete and better organised recall and answered more main-point questions. So decide, before chapter one, who you will explain the book to, and then do it: a partner, a colleague, a notebook pretending to be one. The format is the three-sentence summary, written from memory after the blank page, never copied from the introduction: the claim, the strongest evidence, what it changes. For Smith: "Wealth comes from dividing labour. Ten specialised pin makers produce hundreds or thousands of times what one generalist can, because repetition breeds skill, switching wastes time, and focused workers invent machines. The result is a web of trades so dense that a labourer's coat is the work of thousands." If you cannot write those without the book, you do not yet remember the chapter, whatever it felt like.

Explaining is not using, and the honest answer about using is narrower than the reading advice around it. No experiment on this page tested whether acting on a book makes the book stick. What was tested is which kind of practice reaches the questions that use an idea rather than repeat it, and there the split is clean: the recall advantage in Karpicke and Blunt held on inference questions, while Dunlosky's review finds rereading moving factual questions and finds no effect at all, in Callender and McDaniel's experiments among others, on questions asking for application. So the move the evidence supports is not "apply it" but "practise it in the shape you will need it": if you are reading in order to do something, write the question-notes as decisions ("what do I stop doing on Monday, and why?") rather than as summaries of the text. Choosing books you can act on sits on top of that as a reasonable bet, because a book you use keeps handing you cues to retrieve it with, but that is a bet and not a finding.

Fiction and textbooks

Fiction. A novel's skeleton is characters and turning points. Give each major character a substitute image from the name on first appearance ("Bennet" is a bonnet, "Darcy" a dart) and attach it to the trait the author leads with, the face-name method from the names guide applied to the page; then chain the turning points, the six or eight moments after which things could not go back.

Textbooks. The exam is the retrieval test, so the notes must be in its shape: definitions become question-notes; mechanisms (a pathway, a legal test with four limbs, the causes of a war) become chains, because the examiner wants the order; lists by position go on pegs, a fixed image per number, so the first cause of the war is whatever you did to the peg for one and the fourth is whatever you did to the peg for four.

01major?teaA single steaming cup, a single tea bag, a single stroke. T is one.

A textbook earns the full protocol and a palace, because you already know the month you will need it, and an exam date changes the calendar: the next day, about day 7, about day 20 and a final pass two days before, as in the studying guide, which covers a whole term of this.

The one-book protocol, with time budgets

For a ten-chapter nonfiction book, this adds about fifteen minutes a chapter while reading and about twenty-five more per chapter in reviews spread over the first month.

Then the calendar above: the walk before you put the book down, and the four reviews after it. The only rereading in the protocol is of what you failed at the week-later review.

Questions people ask about remembering what they read

How do you remember a book you read a long time ago?

Sit with a blank page and write what you can, however thin; then read only the contents page and the chapter openings and try again; the contents page is a set of retrieval cues, and the second attempt should return more than the first. Then build the chain or palace you should have built the first time.

How do you connect a book to what you already know and to other books?

Before you read, the connection that helps is the topic itself, which is what the washing-clothes passage shows: context works while you take the passage in, not afterwards. Everything else happens after the recall, on the question-notes. "Why is this true?" is the elaborative interrogation of the notes section above. "What does it contradict that you already believed?" is the question that joins books, because the answer is usually another author and the note should name them both. This page is its own example: Roediger and Karpicke show a second reading losing to a recall attempt, Rawson and Kintsch show a second reading winning when it comes a week later, and the note that holds both, that rereading pays only when it is spaced and never instead of a recall, is easier to keep than either finding alone, because each half cues the other.

Two rules. Join at the blank page, after the recall, so the connection is built on what you retained and not on the open page; and file the notes by theme rather than by book, so a new claim lands beside the ones it argues with and the theme becomes the cue that returns them all.

Does reading faster mean remembering less?

Past a normal pace, yes. Keith Rayner and four colleagues reviewed the reading research in 2016 and found a trade-off between speed and accuracy: a reader is unlikely to go from around 250 words a minute to 500 or 750 and still understand the text as well as at normal speed, and what speed-reading courses and apps mostly amount to is skimming, which is fine when moderate comprehension is the goal. Their route to reading faster with full comprehension is more reading and a bigger vocabulary, because language skill sets reading speed. For a book you want to keep, let the section stop set the pace, and skim only the parts you have already decided not to remember.

Which books are worth reading on this?

Three, for three jobs, and none of them covers all of it. Adler and Van Doren's How to Read a Book is the before: what to do with a contents page, and when a book has earned being put down. Lorayne and Lucas's The Memory Book is the during, the substitute-word drill that turns abstract nouns into pictures, and it is a drill book, so use it as one rather than reading it through. Foer's Moonwalking with Einstein is the why: an account of what competitors actually put in a palace, which will not teach you a system but will tell you what one is for. What all three leave thin is the review calendar, and that is the part deciding whether the rest of it lasts.

Test yourself: ten questions on this article

Close the page, write your answers, then scroll up to check. A reader who only read should, if the lab numbers carry over, get around half; a reader who stopped at each section should get seven or more.

1. On the one-week test, what did a reread passage score, and a recalled one?
2. Five minutes after studying, which was ahead, rereading or recall?
3. In the second concept-mapping experiment, what share of the 120 students predicted concept mapping would be as good as recall or better?
4. What was Ebbinghaus's savings score after one day?
5. What was the passage that only made sense if you knew its topic first?
6. What are the three R's after Survey and Question in SQ3R?
7. Which two study techniques were rated high utility in 2013?
8. What did the story group recall, and the controls, in the twelve-lists experiment?
9. Name the seven kitchen loci, in order.
10. What are the review dates for a chapter read on 14 September?

When you have your score, put your own book's question-notes into a deck and run them through the custom drill, once before you put the book down and once tomorrow. That is the method; the rest was the evidence.