How memory actually works

Why it isn't a file you save and load

Most of us carry around a folk theory of memory that goes something like this: you have a mind, you put things into it, they sit there, and later you take them out again. Like saving and loading a file.

That theory is wrong in interesting ways. The actual architecture of memory (in your brain, in the large language models you may have heard about, and apparently in any system that has to learn under finite resources) is two systems, not one. They work very differently. And the difference matters: it predicts when you'll learn, when you'll forget, and why "I just know" is sometimes a signal of strength and sometimes a signal of trouble.

Two systems, not one

Working memory: the desk you put things on

Right now, while you are reading this, there is a small handful of things active in your mind. Maybe four to seven of them: the sentence you are parsing, the idea you are holding, the name of someone in the next room, the time you need to leave. These are not stored anywhere permanent. They are held by ongoing brain activity: small circuits firing in patterns that keep the information alive. Stop attending, and they fade within seconds.

Cognitive scientists call this working memory. Its defining property is that it is maintained by computation. Nothing is encoded into the substrate; the substrate just keeps re-creating the pattern, moment by moment.

Working memory has a small capacity (4–7 items in humans, varying with content and load). It is flexible (you can rearrange items at will). And critically, it keeps alternatives alive: when you are holding a phone number in mind, you can simultaneously hold the doubt that you remembered it correctly. You know what you do not know.

Long-term memory: the filing cabinet downstairs

If everything had to be held in working memory, you could not function. You could not know your own name, your home address, what apples taste like, how to ride a bike, what last Tuesday was like. The fact that you remember any of those tells you that something else is going on: long-term memory.

Long-term memory is different in kind. It is consolidated into the substrate. Synapses change. Connection weights shift. Physical structure is modified. The information stops being kept alive by ongoing activity and starts being supported by the architecture itself.

Long-term memory has effectively unlimited capacity. It is robust to interruption (forgetting your own name because you got distracted is not a thing that happens). And once consolidated, it is cheap to retrieve. But the cost of those advantages is that the alternatives that were once being weighed during encoding have been compressed away. You do not remember the eleven other candidates your brain considered for "name of my brother". You just remember the answer.

Why "I just know it" is a clue, not a guarantee

This compression of alternatives is what produces the subjective experience of "I just know". When you recall your birthday, you do not feel uncertain; the answer does not come with a hedge. That is because the alternatives that would have produced uncertainty have been pressed away during consolidation. The retrieval surfaces the answer; the architecture has nothing competing to surface alongside it.

Most of the time, that "I just know" feeling is correct. Your birthday really is your birthday. But sometimes (in false memories, in confident wrong answers on an exam, in the everyday phenomenon of fluent confabulation) long-term memory delivers an answer with the same felt certainty whether it is right or wrong. There is no internal alarm bell. The architecture that produces certainty also produces overconfidence.

This is one of the structural reasons why eyewitness testimony is unreliable, why introspection can mislead, and why the most confident speaker in the room is not necessarily the most accurate one. The certainty is a feature of the storage, not of the truth of the content.

You do not learn information. You learn the trace it leaves.

Now for the interesting part. How does information move from working memory to long-term memory? The folk theory says: by paying attention, or by repetition, or by trying hard. There is something to all of those, but they do not explain the mechanism.

Imagine dragging your finger through a thin layer of water on a tiled bathroom floor. The water shifts; a small channel forms. Drag your finger through the same path again, and it is a fraction easier, because the channel is already there. Drag it many times, and you have cut a groove the water naturally follows.

The hysteresis metaphor made literal. The "route" is the path you draw with your finger; the "trace" is the channel the water carves. After enough passes, the channel guides the water and the route can be retrieved with much less effort.

That is hysteresis. The system carries traces of its own history. And it is the precondition for learning. Your brain works exactly the same way. Routes you use a lot leave traces. The traces make those routes more probable next time. After enough use, the route is encoded structurally: you have moved the information from "held by ongoing computation" to "consolidated by substrate change". Learning is that trace-cutting.

This has an uncomfortable implication. You do not learn information. You learn the trace information leaves in you. If you read a textbook passage and your mind does not engage with it (does not try to apply it, does not try to derive consequences from it, does not try to fail and recover), then no trace gets cut. The words wash over the surface. Nothing consolidates.

This is why you can spend hours "studying" and then test poorly. The studying did not fail to give you the information; it failed to make you trace it. And the failure is not a failure of will. It is a failure of physics. The brain does not have a knob you can turn that says "encode harder". Encoding is what happens when routes get reinforced under competition.

The unexpected evidence

Recent evidence from an unexpected place confirms this. Large language models, the neural networks underlying ChatGPT-style systems, turn out to have exactly the same property. You can feed them enormous amounts of information through training, and the resulting model performs worse on the very tasks you trained it for than a model that just has the information in its prompt. The information was given. The trace was not cut. The model does not learn what you fed it; it learns the trace you forced it to make.

The same architectural rule appears in both substrates: they may share the same race-structure, differing in substrate, not in shape. You do not transfer information into a learner. You design conditions under which the learner's substrate cuts the trace.

Why both systems exist

The two-system architecture, working memory and long-term memory as separate regimes, is not an accident or an inefficiency. It is the only architecture that solves a particular trade-off.

You could not live with only working memory. Every fact, every skill, every word of language would have to be held actively, every moment, at metabolic cost. You would run out of capacity in seconds.

You also could not live with only long-term memory. You could not reason about a new situation, hold a tentative hypothesis, or notice that you do not know something. Working memory is what keeps the system honest about its own uncertainty.

Both systems exist because the same trade-off has to be made. Working memory provides calibrated, uncertainty-aware reasoning at high metabolic cost. Long-term memory provides cheap, robust storage at the cost of losing the alternatives. A working architecture needs both. A working architecture is what brains evolved into.

Here is what is genuinely wild about this: no-one designed it. You could not have invented a smarter system for learning (having two regimes with a graceful trade-off between them), and yet no-one sat down and engineered it. The architecture appears in both because they share the same race-structure. The fact that language models, when trained from scratch on a completely different substrate, end up with the same two-regime architecture is the empirical signature that the architecture is structural, not contingent. Brains arrived at it under evolutionary selection. Transformers arrived at it under gradient descent. Neither knew where it was going. Both ended up at the only architecture available.

What this implies for how to actually learn

If learning is trace-cutting, then techniques that cut deeper traces work better than techniques that do not. Educational science has been documenting which is which for decades; the architecture explains why those particular techniques work.

Desirable difficulties (Robert Bjork's term): tests, problems, retrieval practice produce stronger learning than passive re-reading. The reason: difficulty raises route-competition, which deepens the trace. Easy material does not cut a deep groove.

Spacing: spreading study over time beats massed cramming. The reason: each spacing interval lets the trace partly fade, so the next exposure re-deepens it. Cramming just keeps activity high without re-cutting the trace.

Interleaving: mixing different topics forces the brain to discriminate between them, which cuts deeper traces for each. Studying one topic at a time, by contrast, lets routes settle without competing against alternatives.

Active retrieval: forcing yourself to recall something is more effective than re-reading it. The reason: retrieval engages the route, which strengthens it. Re-reading lets the answer surface without effort, which leaves no trace-deepening.

The pattern is consistent: techniques that work all share the property of making the substrate trace the material. The ones that do not work all share the property of not making the substrate do that work. Motivation is mostly irrelevant. Physics is mostly everything.

Technique	What it is	Why it works (trace-cutting mechanism)
Desirable difficulties	Set tasks just hard enough to slow down retrieval	Raises route-competition; the harder the substrate has to work, the deeper the trace
Spacing	Spread study over time instead of cramming	Each interval lets the trace partly fade, so re-exposure re-cuts it; cramming holds activity high without re-cutting
Interleaving	Mix different topics in one session	Forces discrimination between routes; each comparison cuts a deeper trace than either topic alone would
Active retrieval (testing)	Try to recall before re-reading	Retrieval engages and strengthens the route; re-reading lets the answer surface without effort, leaving no deepening

The bigger picture

The architecture of memory is not specifically about brains or about computers. It is about what happens when a system has to learn under finite resources and chooses to do so by leaving traces of its own history. Brains do that. Transformers do that. Some physical systems (magnetisable materials, glasses, certain polymers) carry history-dependent traces too. For brains and transformers the architecture is recognisably the same: working-memory-style states held by computation, long-term-memory-style states consolidated into the substrate, with the trade-offs the same in each substrate. Friction Theory's companion work conjectures that the physical cases may be organisable under the same race-structural vocabulary, a shared way of describing the dynamics rather than a claim that the substrates are identical.

This is not a metaphor. It is, as far as the empirical record can tell, the actual structural reason these systems behave the way they do. The framework that develops the formal version of this is called Friction Theory; the rest of this site explores its implications in language models and in clinical settings, and (more speculatively) whether the same race-structural vocabulary reaches the physics of measurement.