You have met your own memory's cruelest trick, even if you never had a name for it.
You knew the word. You had reviewed it forty times. You could have written it out on request, spelled it, defined it, told me its gender and its plural. And then a man in a bakery in Lisbon asked you a simple question and the word was simply not there — not slow, not fuzzy, not there — and you stood in the warm bread smell feeling like a fraud. Ten minutes later, walking away, it arrived. Unbidden. Perfect. Too late.
Here is the uncomfortable truth: you didn't fail to learn it. You learned it too narrowly. You didn't store a fact. You stored a fact welded to a situation — a white screen, a particular font, a particular question phrased a particular way, your bedroom at 11pm — and when the situation changed, the weld snapped. Memory researchers have a name for what you were missing, and it is one of the most useful ideas in cognitive psychology that almost nobody outside the field talks about: encoding variability.
A memory is not a thing. It is a set of doors.
We talk about memory as storage. Facts as files. Forgetting as deletion. Every one of those metaphors is wrong in the same direction, and the wrongness costs you real hours.
A better picture, and the one the research actually supports, is this: a memory is a trace bound to the cues that were present when you formed it. Endel Tulving and Donald Thomson formalized this as the encoding specificity principle — what makes an effective retrieval cue depends on what was encoded alongside the memory in the first place. The information doesn't sit in a room labeled with its meaning. It sits behind a door, and the door only opens for the keys that were in the lock when you closed it.
Study a word once, in one form, in one place, and you build one door. It is a beautiful, well-oiled door. It opens instantly — when you approach from that exact hallway.
The bakery in Lisbon is a different hallway.
Why repetition alone quietly betrays you
Now the twist. Reviewing the same card, in the same way, forty times does not build forty doors. It mostly polishes the one.
This is the core of the encoding variability hypothesis, proposed by Edwin Martin and developed by Arthur Glenberg and others in the 1970s. The claim is that repetitions help most when the context of each repetition differs — when the fluctuating internal and external background at study time isn't identical each round. Each varied encounter attaches the memory to a slightly different set of cues. Over time you're not deepening one channel; you're widening the delta of channels that can carry you to the fact.
The cleanest demonstration is an old and famous one. Smith, Glenberg, and Bjork had people study material in either a single room across two sessions or in two different rooms, then tested them in a third, neutral place. The two-room group recalled more. Nothing about the material changed. Nothing about the total study time changed. What changed was the number of distinct contexts the memory was tied to — so that later, in a place matching neither, some cue was still in reach.
This is also, quietly, part of why spacing works at all. When you review something a week later, you are a slightly different person in a slightly different mood in a slightly different room. Spacing doesn't just let forgetting do its useful work; it forces variability on you. It is variability with a schedule.
Cue overload: the other half of the story
There's a companion mechanism that explains the bakery specifically, and it is worth naming because it is counterintuitive.
Cues compete. This is the cue overload principle: the more items a single cue points to, the worse it points to any one of them. If "the front of this card" is the only key you've ever used, that key is doing all the work — and it's a key you'll never have in your hand during an actual conversation.
Worse, single-context learning produces exactly the fluency that fools you. You feel fast. You feel certain. That feeling is generated by the cue, not by the knowledge. Robert Bjork's distinction between storage strength (how well-learned something is) and retrieval strength (how accessible it is right now) is the sharp edge here: massed identical repetition inflates retrieval strength in the moment while doing comparatively little for the durable, flexible representation you actually wanted. You are paying for the feeling of knowing and receiving very little of the knowing.
The fix is not more reps. The fix is different reps.
What variability actually looks like
This is where people go wrong in a well-intentioned way. Encoding variability is not "study in a café one day and a library the next" — that's the shallowest possible reading, and the effects of physical room context are real but modest and mostly matter when everything else is held constant.
The variability that pays is variability in how the information is engaged, because that's what changes the internal cues, the ones that travel with you:
- The same fact approached from a different direction (given the answer, produce the question).
- The same fact expressed in a different modality (spoken aloud, sketched, gestured).
- The same fact embedded in a sentence rather than floating alone.
- The same fact connected to something you already know — a memory, a person, a place.
- The same fact tested with a different prompt than the one you studied.
Each of these hangs a new door on the same room. And crucially, it will feel worse. Varied practice is slower and more error-prone during learning than blocked, identical repetition — and it produces better retention and better transfer afterward. This is what Bjork means by a desirable difficulty: the conditions that make performance look best today are frequently not the ones that make learning last.
You have to be willing to look worse in practice to be better in the bakery.
Your next moves
Do these today, on material you're already studying.
- Take your five most-reviewed cards — the ones you always get right — and reverse them. Make a card that shows the answer and asks for the term. If the reverse feels hard while the forward feels trivial, you've just proven you had one door, not knowledge.
- Rewrite the prompt on three cards you keep failing. Don't add information. Change the phrasing of the question, or turn the bare word into a sentence with a blank in it. You are trying to learn the fact, not the question.
- Say one deck out loud in a different room from where you made it, once, standing up. Speaking recruits articulatory and auditory cues that a silent screen never touches — the production effect and encoding variability compound here.
- For every abstract fact, write one line connecting it to a specific person, place, or moment in your life. "Saudade — what my grandmother's kitchen felt like the year after she died." That line is a permanent second key, and it's one nobody can take from you.
- Deliberately review at a bad time. Tired, distracted, on a train. Not always — but occasionally. Retrieval that succeeds in poor conditions is retrieval that isn't leaning on a pristine, unrepeatable context.
The one thing a review app should be doing for you
Everything above is a way of saying: a fact is only as retrievable as the number of honest, different roads that lead to it. Spaced repetition earns most of its power not because the algorithm knows some secret about your brain, but because it drags you back to the same idea across genuinely different versions of yourself — different days, moods, rooms, states of forgetting — and each return trip lays another road.
That's the part we tried to build Recall around: a review schedule (modern FSRS) that spaces you out far enough for real variability to accumulate, cards that are quick enough to reverse and rewrite that you actually will, and a deck that lives in your pocket and works fully offline, so a review can happen on the train, half-asleep, in the wrong room, under the wrong conditions — which, it turns out, are the right ones. Bring your Anki and Quizlet decks over and start widening the doors.