Intelligence without memory. The doctor who forgets your case on closing the door

An LLM (large language model) doesn't remember what you talked about yesterday. Every conversation is born virgin, held up only by what fits in its context window, a strip of text that, in the models of the mid-decade, ran from the few thousand tokens of the cheap models to the two million Gemini 1.5 Pro went so far as to announce. Outside that strip, for the model, nothing exists. If you went to a doctor who forgot your case on closing the consulting-room door, you'd say they weren't a good doctor. We accept it from AI because the other option is realising it's an assistant with amnesia. And that would force us to review half the products we've already wired into our lives.

The doctor metaphor, seriously

Picture a consultation. You describe your symptoms. The doctor listens, asks questions, writes a brilliant report while you're in front of them. You close the door. You come back the next day. The doctor receives you with the same professional courtesy, with no trace of who you are or what you discussed yesterday. You explain the pain again, the tests, the context. They're a good doctor, within each visit. Between visits you're nothing.

That's what an LLM is. And yet we call it intelligent with the same word we use to call the psychotherapist who has followed your case for fifteen years intelligent. It isn't an imprecision of language. It's an ontological difference disguised as a technical nuance, and the nuance suits whoever sells the product.

What the context window is and why it isn't memory

The minimum unit a model processes is called a token: roughly half a word in English, a bit less in Spanish. The number of tokens a model can have in front of it simultaneously, reading and predicting, is its context window. GPT-3 worked with around two thousand. GPT-4 expanded to eight thousand, then to thirty-two thousand, and GPT-4 Turbo reached one hundred and twenty-eight thousand. The first versions of Claude, the 2023 ones, ranged between one hundred thousand and two hundred thousand tokens, and later models went as far as announcing a million. Gemini 1.5 Pro boasted of two million. Any specific figure ages in months, and by the time you read this the record will be another; what matters is that the window grows and the structure doesn't change.

Within the window, the model can operate on everything that's there. Outside the window, the model is blind. There's no raw archive it can go back to, no earlier episode consolidated anywhere. Yesterday's conversation, if it exceeds the window's size, has vanished for it the same way page 78 of the book you haven't opened yet has vanished for you. Only that for you it exists in reality and can be consulted; for the model it doesn't exist.

The context window is working memory in the strict sense. It's what, in a human brain, would resemble the phonological buffer: what you're actively holding while you process. It isn't long-term memory. And the difference between the two isn't quantitative, it's qualitative: they're different systems, with different neurochemistries, with different functions. Confusing them has been, for thirty years, the standard pedagogical error of popular neuroscience. Confusing them in the case of LLMs is, on top of that, lucrative.

How the memory that isn't there gets faked

There are three mechanisms the industry packages under the word "memory", and none of them is.

The first is the conversation buffer. The system wrapping the model, what the user calls "the ChatGPT" but which technically is a layer over the model, saves what you said before and feeds it back into the window each turn. That's why it seems the chatbot remembers what you asked three messages back. It doesn't remember it. It's being re-fed to it, disguised, in each inference. When the conversation exceeds the window's size, the oldest messages fall over the edge like a sock down the drain. Some systems summarise them before throwing them out, which generates a memory by compression that loses detail at every step. Remembering and having a progressively degraded summary of what happened are not the same.

The second is RAG, retrieval-augmented generation. The technique, formalised by Lewis and others at NeurIPS 2020, works like this. When you ask a question, an external system searches a vector database for the text fragments most similar to your query and pastes them in front of your question, inside the window, before passing it all to the model. It's indexing, not remembering. The model doesn't "know" what's in the database. It has no experiential access to that past. It receives, at the last moment, a dossier of fragments a search engine has selected by cosine similarity, and it reads them as it would read a PDF you'd just attached. The right metaphor isn't a doctor's memory, it's the intern who passes the doctor your file just before you come through the door.

The third, more recent, are the persistent agents: customised GPTs, Claude Projects, assistants with "user memory". The name sounds like something else but the mechanism is the same as the first case, refined. The system extracts meaningful sentences from earlier conversations, stores them externally, and re-injects them into the window when it estimates they're relevant. The selection is done by a process that, logically, is also a model, with its own failures. If the system decides something isn't relevant, that stays out. If it decides something is and gets it wrong, that goes in and skews the answer. The "memory" of the customised assistant is an opaque decisional box whose rules the user doesn't control and usually doesn't know.

Anterograde amnesia as the default condition

There's a clinical name for what an LLM suffers from. It's called anterograde amnesia. The patient can't form new long-term memories after the brain damage, even though they keep the earlier ones perfectly and their working memory functions within each interaction.

The paradigmatic case in the literature is H. M., Henry Molaison, the patient on whom Scoville and Milner published in 1957 the famous Loss of recent memory after bilateral hippocampal lesions in the Journal of Neurology, Neurosurgery, and Psychiatry. H. M. had had his hippocampus bilaterally removed to treat drug-resistant epilepsy. The epilepsy improved. What was left was a man who, for fifty-five years, didn't consolidate a single new memory. He held sophisticated conversations. He learned motor skills, which showed that procedural memory uses other circuits. But the next day, the researcher who'd been studying him for decades he greeted as a stranger.

Clive Wearing and the thirty-second consciousness

The other case is Clive Wearing, a British musician who in 1985 suffered a herpetic encephalitis that destroyed his hippocampus and parts of his temporal cortex. His wife Deborah wrote a whole book about what it's like to live with someone whose window of consciousness lasts between seven and thirty seconds, and who spends every day rediscovering, noting in a notebook "NOW I am awake for the first time", only to cross out the line and rewrite it ten minutes later on losing the thread. The book is hard to read. It's the closest thing to understanding, from outside, what it is to operate with a working memory disconnected from consolidation.

An LLM does structurally the same. The difference is that an LLM doesn't suffer, as far as we know. And that the damage isn't damage, it's architecture. But the cognitive operation, read from outside, is that of a brain with a destroyed hippocampus whose day gets reset every time it crosses a door.

When someone argues that calling this intelligence is a harmless hyperbole, the reasonable reply is to ask whether they'd call Clive Wearing operating a banking system intelligence.

RAG doesn't consolidate. It indexes

It's worth insisting, because the confusion is deliberately sustained by the selling side. Tulving in 1972 distinguished between episodic memory (the facts you lived with a temporal and emotional marker, "I remember when") and semantic memory (the facts you know with no autobiographical context, "Paris is the capital of France"). Human memory works because it has both, connected, and because it can transfer information from one to the other through consolidation, which is a slow, biological process, costly in terms of sleep and neurogenesis. Schacter, in The Seven Sins of Memory (Houghton Mifflin, 2001), described how far that consolidation is, on top of that, imperfect and reconstructive: each time you remember, you rewrite.

What RAG offers fits in neither of the two. It's an externalised semantic search memory, a stack of fragments accessed by similarity and injected into the window without passing through anything resembling consolidation. The model doesn't integrate those fragments into its structure. It doesn't weigh them against others. It doesn't rewrite them in the light of later experiences. It processes them once, in the current turn, and drops them when the conversation closes. If you come back tomorrow, the system searches again, retrieves other fragments, and the model gets different ones.

Calling this memory is like calling having access to Google memory. Access to Google is great. It isn't what we used to do when we said someone remembered something.

The assistant with memory, the customer with expectation

The advertising trick is old and effective. The industry takes a word from the human register, memory, sticks it on top of a technical mechanism that has little to do with it, and lets the user fill the gap with their own experience. The user has memory, has spent their whole life with it, knows what it is and what it feels like to remember. When they read "the assistant remembers what you've told it", they assume the term's full sense. And what happens technically is what was described above: a process that extracts sentences, indexes them, and re-injects them as additional context when another process decides so.

The difference between one thing and the other seems a matter of nuance. It isn't. If a human psychotherapist follows your case, what happens when you come back is that they remember who you are, with its emotional charge, with their private hierarchy of what was important and what was noise, with the narrative they've been weaving over months. When you come back to a customised GPT, what happens is that a process has decided which three or four facts about you were salient enough to keep, and gives them back to you as bullets in the system prompt. It isn't a nuance: they're two different systems serving the same function at the price of defining the function differently.

Gary Marcus has spent years complaining, rightly, about the elastic use of the term. It's one of the very few things on which it's worth granting him the point without picking a fight with the rest of his discourse first. Bender, Gebru, McMillan-Major and Shmitchell laid out the general case in On the Dangers of Stochastic Parrots (2021): every time the industry imports human cognitive vocabulary to describe a computational mechanism, it transfers free to the product the affective charge the word drags along. "Memory" is the most profitable piece of that transfer.

The difference marketing needs to hide

There's a question the industry doesn't want asked aloud. If an intelligence without memory is something structurally different from human intelligence, what are we measuring when we compare its performance to a human expert's on a cognitive task?

The honest answer is that we compare point results on bounded tasks within a window. That the model shines in an exam doesn't mean it can carry a case over a year. That it passes a moral-reasoning test doesn't mean it'll hold the same criterion tomorrow. That it answers a medical query perfectly doesn't mean it can follow the patient up. What intelligence without memory can do is solve the episode. What it can't do is accompany the case. And most serious human intellectual work, outside exams, is accompanying the case.

The generative agents of Park and others (arXiv 2304.03442, 2023) are the most serious academic attempt to bridge this ditch. They build a simulated memory architecture with three levels: observation, reflection, planning. Episodic memory is stored as a chain of scorable events, the most relevant are retrieved by importance, recency and similarity, and a periodic reflection process generates high-level summaries that are also saved. The result is impressive in demos. It works in simulations of digital towns. It doesn't yet work outside those closed environments where the computational cost would let it scale to real human scale. And, most importantly, it's still text indexing, not neural consolidation. It's an intelligent substitute. It isn't the thing.

When Kandel won the 2000 Nobel for discovering the molecular mechanisms of long-term memory in Aplysia californica, and expanded them in In Search of Memory (Norton, 2006), he was describing something that happens by modification of synapses, by protein synthesis, by irreversible structural changes in the neurons involved. That's consolidating. That's what makes the psychotherapist from before still remember you. It has nothing to do with computing cosine distances between vectors.

The honest conversation starts by accepting this

Accepting that an intelligence without memory is structurally different from ours resolves nothing. It doesn't say whether it's worse, doesn't say whether it's better, doesn't say whether it'll end up compensating with brute power for what it lacks by architecture. It only says that calling it by the same word is a choice, and that the choice benefits someone.

An honest conversation about what an LLM can and can't do, about where to delegate to it and where not, about which economic sectors it'll transform and which it'll only erode leaving a hole, can't begin by pretending the system remembers. It has to begin by saying it forgets everything every time you close the door, and that each interaction is a first consultation. That changes, in one stroke, the list of professions the system can replace, the ones it can only assist, and the ones that rest precisely on the property the system doesn't have.

Most of the commercial texts about AI doing the rounds today haven't got round to accepting this. And that's why, for the most part, they aren't texts about AI. They're texts about how AI resembles us if we look at it in the right light.

Definitions

Token. The minimum unit of text a language model processes. Equivalent to roughly half a word in English, slightly less in Spanish. Models don't read letters or words: they read sequences of tokens.

Context window. The maximum number of tokens a model can have in front of it simultaneously when generating a response. It works as working memory, not as long-term memory. Everything exceeding the window stays out of the model's reach.

RAG (retrieval-augmented generation). A technique combining a language model with an external database. When a query arrives, a system searches the database for relevant fragments, injects them into the model's window and lets it answer supported on that material. It isn't memory; it's indexing with just-in-time delivery.

Embedding. A numerical representation of a text fragment as a vector in a high-dimensional space. It allows similarity between texts to be computed by geometric distance, and it's the technical basis of the search RAG runs over the external database.

Anterograde amnesia. The inability to form new long-term memories after brain damage, usually keeping the earlier ones and the moment's working memory. It characterises structurally patients with hippocampal destruction and, by operational analogy, LLMs by default.

Episodic memory. The recall of lived facts with a temporal marker and subjective charge. It's the memory that lets you say "I remember when". Tulving distinguished it in 1972 from semantic memory.

Semantic memory. Knowledge of facts with no autobiographical context. Knowing that Paris is the capital of France without remembering when you learned it.

Consolidation. The biological process by which a working memory transforms into long-term memory, through synaptic changes and protein synthesis. Kandel described the main molecular mechanisms. It's what LLMs, by architecture, don't do.

Persistent agent. A system built on top of an LLM that stores user information externally and re-injects it into the window when it considers it relevant. Customised GPTs and Claude Projects are current commercial examples.

References

Lewis, P., Perez, E. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. The formal origin of RAG, cited to distinguish indexing from memory.

Park, J. S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv 2304.03442. A simulated memory architecture in three levels, cited as the most serious academic attempt to close the persistence ditch.

Tulving, E. (1972). Episodic and Semantic Memory. In Organization of Memory, Academic Press. The framework distinguishing episodic from semantic memory, the basis of the argument on why RAG isn't memory.

Schacter, D. L. (2001). The Seven Sins of Memory. Houghton Mifflin. Cited to underline that human consolidation is, on top of that, reconstructive and imperfect, not a clean archive.

Scoville, W. B. & Milner, B. (1957). Loss of recent memory after bilateral hippocampal lesions. Journal of Neurology, Neurosurgery, and Psychiatry, 20, 11–21. The H. M. case, the clinical paradigm of anterograde amnesia, cited for the structural analogy with the LLM's default condition.

Wearing, D. (2005). Forever Today: A Memoir of Love and Amnesia. Doubleday. A front-line testimony on life with a conscious window of seconds, cited for its analogical value.

Kandel, E. (2006). In Search of Memory: The Emergence of a New Science of Mind. W. W. Norton. The molecular mechanisms of long-term consolidation, cited to pin down exactly what an LLM doesn't do.

Bender, E., Gebru, T., McMillan-Major, A. & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT 2021. DOI: . Cited for the general critique of the elastic use of cognitive terms applied to language models.

Google. Long context. Gemini API documentation. The source for the figure of up to two million context-window tokens in Gemini 1.5 Pro.

Marcus, G. Marcus on AI. Substack. A continued critique of the use of the term "memory" applied to buffers and to RAG.

To go deeper

Sacks, O. (1985). The Man Who Mistook His Wife for a Hat. And Other Clinical Tales. Summit Books. A gallery of clinical cases on memory dissociations that support, from neurology, the idea that "memory" names many things and not one.

You might also like

• Learning without remembering. The word the industry abuses
• The context problem. The invisible cage of every human-AI conversation
• Types of memory
• Thinking without understanding. If it works the same, what did we think thinking was

Elsewhere

• Henry Molaison (Wikipedia): patient H. M. and anterograde amnesia
• AI: A Guide for Thinking Humans (Melanie Mitchell)