RAM memory vs human memory. Two different things with the same word

In 1974, Elizabeth Loftus showed a group of people a recording of a car crash. Some she asked how fast the cars were going when they "smashed." Others, how fast when they "contacted." The first gave higher figures. A week later, they additionally remembered broken glass that had never appeared in the video. Human memory doesn't store: it rewrites. Every time you remember, you edit. A computer's memory doesn't do that. It returns what it saved, identical, and wipes when you switch it off. Calling them by the same word is a linguistic loan that hides a difference of nature, not of degree.

You have memory. Computers, big or small, have something else. It's called the same, but the two aren't remotely alike. The word is the same because it struck someone in the fifties as convenient; the mechanism, not. Worth keeping in mind from the first line, because almost all the public noise about artificial intelligence lives in that confusion.

What RAM does

RAM (Random Access Memory: the fast, temporary workspace of a computer) is a grid of cells that hold ones and zeros. Each cell has a numerical address. You ask the system for cell number 4,812,317 and it returns it, intact, in nanoseconds.

If you ask for it a thousand times in a row, a thousand times it returns the same thing. If you turn the computer off, everything disappears without a trace.

That's what RAM does and it does nothing else. It's fast, exact and volatile. It doesn't interpret, doesn't prioritize, doesn't decide what's worth keeping. It's a shelf with numbered slots and a very diligent stock clerk.

The quality metric is simple. How many gigabytes it holds, in how many nanoseconds it responds, how many cycles per second it handles. You buy better RAM and you get more of the same, faster. The scale is linear and the function is stable. That, at least, is clean.

What your brain does

Your memory doesn't have numbered cells. Nor does it have a bit-by-bit copy mechanism. What you have is a network of neurons that have been connecting over the course of your life, where the trace of a memory is a pattern spread across thousands of synapses that strengthen or weaken according to a logic we still don't fully understand.

When you remember, you don't open a file. You reconstruct.

Each recall is a small act of re-editing. Your current state, your emotion right now, what you've just been told, what you expect to happen tomorrow, all of it mixes with the original trace and modifies it. Human memory is a process, not a warehouse.

Three memories instead of one

Endel Tulving, in 1972, proposed a distinction that has stuck. Episodic memory stores lived scenes with their temporal and emotional context: your wedding, your first bicycle, the day your father died. Semantic memory stores facts without personal context: that Paris is the capital of France, that water boils at a hundred degrees. Tulving himself, in 1985, expanded the scheme to explicitly incorporate procedural memory: the one that stores motor skills and automatisms —riding a bike, typing, tying a knot— and runs without passing through consciousness.

They're distinct systems, housed in distinct brain structures, that can be damaged independently. There are patients who lose personal memories and keep their knowledge of the world intact. There are others who keep the bicycle in their legs and have forgotten who they are. The plurality isn't metaphor: it's anatomy.

RAM has none of this. RAM is flat. A neuron is plural.

The tape that rewrites itself

The Loftus experiment that opens the article is one of the most replicated in cognitive psychology. Changing a verb in the question changes the memory. The word "smash" introduces violence, and the violence introduces broken glass that never existed. The memory reorganizes itself to sustain the coherence the question suggests.

Loftus then spent decades demonstrating something even more uncomfortable. Whole memories can be implanted. In a study known as the Disneyland one, her collaborators convinced a significant share of participants that as children they had shaken hands with Bugs Bunny at the park. Bugs Bunny is Warner's. He was never at Disneyland. The memory was new, false, and completely vivid for whoever held it.

This isn't a laboratory oddity. It's how your memory works every day.

To consolidate is to re-edit

Consolidation (the process by which a short-term memory gets fixed in the brain over the hours and days that follow, above all during sleep) isn't a recording. It's a re-edit. And every time you reactivate a memory, you reopen that editing window. What you remember today of your childhood isn't what you lived as a child. It's the latest revised version of a text you've been rewriting for forty years.

James McGaugh and others further showed that emotion marks that editing with a special seal. Emotionally charged memories consolidate more strongly and reactivate more easily, thanks to the amygdala's intervention on the hippocampus. We remember better what frightened us, what made us fall in love, what hurt us. And by remembering it better, we deform it more, because we summon it more often.

Why your "worse" memory is better

Seen through a computer scientist's eyes, this system is a disaster. It rewrites data without permission. It loses information all the time. It's slow. It makes mistakes. It has no audit log. Any database that worked like this would be pulled from the market the next day.

And yet, it's what lets you be you.

Your brain doesn't need to keep everything. It needs to keep what serves to survive, decide, bond and narrate yourself. That means remembering the face of the one who betrayed you and forgetting what you had for dinner on Tuesday. It means remembering the smell of your grandmother's house and forgetting the name of the classmate who switched schools in third grade. It means a trauma reactivating thirty years later at a similar noise, because to your brain that information still seems relevant to keeping you alive.

Daniel Schacter called the failures of human memory "the seven sins" and showed that each one is the visible face of a useful mechanism. We quickly forget the irrelevant because retaining it would saturate us. We accept suggestions in our memories because social integration weighs more than photographic fidelity. We bias what we remember toward our self-image because without a coherent self-image we couldn't function.

The two extremes. Remembering everything or fixing nothing

There's a brutal counterexample. Solomon Shereshevsky, the mnemonist the Soviet neuropsychologist Aleksandr Luria studied for thirty years, was incapable of forgetting. He remembered meaningless tables of numbers read once, fifteen years earlier, in exact order. And he was a wretched man.

He couldn't abstract. He couldn't read a text without every word dragging along an avalanche of specific images. He couldn't hold a normal conversation because each sentence triggered a thousand associations he couldn't switch off. Remembering everything is a way of not living.

Anterograde amnesia (the inability to fix new memories after a brain injury, as happened to the patient known as H.M. after the removal of his hippocampi in 1953) shows the other extreme. If you stop fixing the new, there's no subject left either. Human memory works because it selects, prioritizes and deforms. It's not a bug. It's the function.

The trap of the metaphor

When someone says the brain holds "the equivalent of two hundred petabytes" or that a language model "has a better memory than a human," they're committing a category error. They're measuring two different things with the same ruler because they share a label. It's like comparing the speed of a plane with the speed of light because both are measured in kilometers per second.

RAM and your memory don't solve the same problem. RAM solves the problem of having exact bits available in little time for a processor to work with them. Your memory solves the problem of building a continuous self in a body that has to make decisions in a changing world. They're different functions, evolved or designed on different planes, subjected to different pressures.

Comparing their capacities is like comparing the capacity of a hammer with the capacity of a conversation. Both can be expressed in some number if you insist. None of those numbers means anything.

And language models, what do they have?

A large language model has three things we tend to call memory, and it's worth separating them.

It has the model's weights. A huge matrix of numbers learned during training, encoding statistical regularities of the text it was trained on. This is not episodic memory of anything. It's a statistical distillate with no scenes, no temporal context, no subject doing the remembering.

It has the context window (the maximum amount of text the model can look at at once when generating a response, usually tens or hundreds of thousands of words). This is closer to a human's working memory, but without functional forgetting or consolidation. When the conversation closes, it vanishes like RAM at shutdown.

And it has external mechanisms. Embeddings (numerical representations of words or phrases in a mathematical space that allows meanings to be compared by closeness) and the technique known as RAG (Retrieval Augmented Generation: the model searches for text fragments in an external database and injects them into the context window before answering). The system breaks documents into pieces via chunking (splitting text into fragments the right size to index them), turns them into vectors, looks them up when needed and pastes them into the input. It's a librarian, it's not a memory.

Why none of this remembers

None of the three things does what your memory does. None deforms with emotion. None consolidates in sleep. None forgets the irrelevant to protect the coherence of the self. None builds a continuous identity. The three are information-retrieval architectures, optimized to return relevant fragments. That's valuable. But it isn't memory in the human sense. It's something else.

When a model "remembers" that in an earlier conversation you told it your name, it's reading an external file where someone saved it. When it "remembers" a historical fact, it's reactivating a statistical pattern in its weights. At no point is there a subject who lives the scene of knowing it and to whom that knowing matters.

The difference that doesn't close with more gigabytes

You might think the difference is a matter of scale. More parameters, more context, more databases, better retrieval algorithms, and in the end the system will resemble it closely enough. It's a reasonable intuition and it's probably false.

Human memory isn't information stored with variable efficiency. It's a mechanism built by evolution under the pressure of keeping an organism alive and a subject coherent. Each of its technical "failures" is a response to that pressure. Forgetting isn't storage savings: it's protection of the self. Rewriting isn't an integrity bug: it's the updating of a model of the world that has to keep working.

For a machine to have human memory, storing everything better wouldn't be enough. You'd have to endow it with a body that could die, a history that could go wrong, a future that mattered. Without that, what you have is an archive. The archive can be immense, can be lightning-fast, can be astonishingly useful. It isn't memory. It's something else.

RAM doesn't remember: it returns what it holds. Your brain does remember, because you have memory. Computers have something else, and as long as we keep calling by the same word two mechanisms that share neither the function nor the subject, we'll keep expecting from machines something they can't give, and distrusting ourselves for not resembling them.

Definiciones

RAM (Random Access Memory). The fast, volatile work area of a computer. It stores bits in cells with numerical addresses, returns them identical each time they're requested, and wipes when the power is cut.

Episodic memory. A memory system that stores lived scenes with their temporal and emotional context. It's the memory of what happened to you, not of what you know.

Semantic memory. A memory system that stores general knowledge without personal context: facts, concepts, word meanings. You know Paris is the capital of France, but you don't remember when you learned it.

Procedural memory. A memory system that stores motor skills and automated procedures, like riding a bicycle or driving. It runs without passing through consciousness.

Anterograde amnesia. The inability to fix new memories after a brain injury. The patient keeps what they knew before, but can't form memories afterward.

Consolidation. The process by which a short-term memory becomes durably fixed in the brain over the following hours and days, largely during sleep. Consolidation re-edits the memory each time it's reactivated.

Embeddings. Numerical representations of words or phrases in a high-dimensional mathematical space, where geometric proximity corresponds to closeness of meaning. Used to search for "similar" text in automated systems.

Context window. The maximum amount of text a language model can look at at once when generating a response. It works like a temporary working memory: what leaves the window stops existing for the model.

Chunking. The technique of splitting a document into fragments the right size to index them in a vector database and retrieve them by similarity.

RAG (Retrieval Augmented Generation). An architecture in which a language model queries an external database before answering, retrieves relevant fragments and injects them into the context window. It isn't memory: it's an attached librarian.

Referencias

Loftus, E. F. & Palmer, J. C. (1974). Reconstruction of Automobile Destruction: An Example of the Interaction Between Language and Memory, in Journal of Verbal Learning and Verbal Behavior 13, 585–589. Source of the car experiment and of the introduction of broken glass into the participants' memory.

Loftus, E. F. (2003). Make-Believe Memories, in American Psychologist 58, 867–873. A synthesis of the work on experimental implantation of false memories, including the study known as the Disneyland one cited in the article. Complementary reference at Wikipedia — Elizabeth Loftus.

Tulving, E. (1972). Episodic and Semantic Memory, in Organization of Memory, ed. E. Tulving and W. Donaldson, Academic Press. The foundational distinction between episodic and semantic memory used in the article.

Tulving, E. (1985). How Many Memory Systems Are There?, in American Psychologist 40, 385–398. A later review in which Tulving himself explicitly incorporates procedural memory into the scheme.

McGaugh, J. L. (2003). Memory and Emotion: The Making of Lasting Memories, Columbia University Press. Source of the argument on the emotional mark in consolidation and the role of the amygdala.

Schacter, D. L. (2001). The Seven Sins of Memory: How the Mind Forgets and Remembers, Houghton Mifflin. Origin of the framework that presents the functional failures of human memory as the counterpart of useful mechanisms.

Luria, A. R. (1968). The Mind of a Mnemonist: A Little Book About a Vast Memory, Basic Books. The classic study of Solomon Shereshevsky, the mnemonist who couldn't forget, used in the article as a counterexample.

Para profundizar

Kandel, E. (2006). In Search of Memory. The Emergence of a New Science of Mind. Norton. An intellectual biography of the neural study of memory by one of the Nobel laureates who contributed most to the field.

Damasio, A. (1999). The Feeling of What Happens. Body and Emotion in the Making of Consciousness. Harcourt. The connection between emotion, body and memory as layers of consciousness, useful for understanding why the substrate of memory doesn't separate from the organism.

Loftus, E. F. (1979). Eyewitness Testimony. Harvard University Press. The judicial application of the reconstructive properties of human memory.

También te interesa

En otros sitios

#ai-memory #anthropomorphism #intelligence #papers