AI’s Memory Problem Was Discovered in an Operating Room 70 Years Ago
Why does AI keep getting smarter — yet still struggle to remember you?
The Man Who Could No Longer Form New Memories
In 1953, a 27-year-old man lay down on an operating table in Connecticut.
His name was Henry Molaison, though the neuroscience community would know him for decades only by his initials: H.M. He had suffered from severe epilepsy since childhood, and medication could no longer control his seizures. His doctors proposed a radical treatment: removing portions of both medial temporal lobes, including a structure whose function was still poorly understood — the hippocampus.
The surgery was a “success.” His seizures improved.
But the cost was extraordinary.
H.M.’s intelligence remained intact. His personality did not change. He could carry on a perfectly normal conversation, and he still remembered events from his childhood. But from the day of the surgery forward, he could no longer form lasting memories of new experiences.
Each time a doctor entered his room, H.M. greeted them as if they had never met. He could talk with you happily, but if you stepped out for more than a minute and returned, the conversation was gone. In a sense, he was permanently stranded in 1953.
Then researchers discovered something even stranger.
They asked H.M. to practice a difficult mirror-drawing task: tracing a star while looking only at its reflection. His performance improved day after day — even though every day, he insisted he had never done the exercise before.
He could not remember what had happened, but he could still learn how to do something.
The finding transformed cognitive science because it demonstrated, for the first time, a fundamental truth:
Memory is not one thing. It is a collection of independent systems located in different parts of the brain — and those systems can fail separately.
Seventy years later, anyone who has used ChatGPT or another AI assistant has probably experienced a strangely similar situation.
You spend two hours explaining your project, your preferences, and all the mistakes you have already made. The next day, you open a new chat and it cheerfully asks, “How can I help you?”
Today’s AI systems are all H.M.
And some of the world’s brightest engineers are now doing the same thing: studying the architecture of the human brain and adding memory to AI, layer by layer.
Your Brain Isn’t a Hard Drive. It’s an Orchestra.
Our intuitive model of memory is simple: the brain is a hard drive. Information goes in, gets stored, and is retrieved when needed.
More than six decades of cognitive science tell us otherwise.
In 1968, psychologists Richard Atkinson and Richard Shiffrin proposed the multi-store model of memory. In their framework, information flows through three gates: sensory memory → short-term memory → long-term memory.
In 1972, Endel Tulving divided long-term memory into two categories: episodic memory — remembering what happened — and semantic memory — knowing what something is. Researchers later added procedural memory — knowing how to do something.
Neuroscience has since confirmed that these forms of memory rely on different brain regions, each with its own mechanisms for encoding, storage, and retrieval. Different dimensions of the same experience can exist independently.
That is why H.M. could not remember practicing the drawing task the day before, yet continued to improve. The hippocampal system involved in remembering events had been damaged, while the brain systems supporting skill learning remained largely intact.
For the purpose of understanding AI, we can organize these systems into a seven-layer map. Remarkably, each layer has a counterpart that today’s AI products are attempting to replicate:
- Sensory memory: seeing and hearing
- Short-term memory: briefly retaining information
- Working memory: understanding and processing
- Episodic memory: remembering an experience
- Semantic memory: turning experience into knowledge
- Emotional memory: deciding what matters
- Procedural memory: turning experience into skill
To make the differences concrete, let’s look at something ordinary: reading a book.
What Happens in Your Brain When You Open a Book
Layer 1: Seeing — Sensory Memory
As your eyes move across a page, the visual cortex in your occipital lobe instantly captures the shapes of letters, line spacing, and layout.
This is sensory memory. Visual sensory memory lasts only about half a second; auditory sensory memory persists slightly longer, for roughly three to four seconds.
Its capacity is enormous, but it decays almost immediately — like a beach constantly being washed clean by the tide. Only information selected by attention reaches the next stage.
While you focus on a sentence on page 42, the page number in the corner and the shadows of trees outside your window also enter sensory memory. Within half a second, most of that information is discarded.
How AI copies it
This is AI’s perception layer. Optical character recognition turns images into text. Vision-language models such as GPT-4V and Qwen-VL interpret visual layouts. Whisper converts speech into text.
Without this layer, AI is both blind and deaf. The equivalent of selective attention is information preprocessing: not every input deserves to be remembered. The system first has to identify what matters.
Layer 2: Holding — Short-Term Memory
As you continue reading, the current sentence remains in your mind just long enough for you to connect its second half to its first.
This is short-term memory, which typically holds information for about 15 to 30 seconds.
Its capacity is notoriously limited. In 1956, psychologist George Miller proposed the famous “seven, plus or minus two” rule: people can maintain roughly five to nine chunks of information in short-term memory at once. That is why we break long phone numbers into smaller groups.
How AI copies it
The closest equivalent is a large language model’s context window. Everything in the current conversation lives there. Once it falls outside the window — or the session ends — it may effectively disappear.
That is the immediate reason AI appears to “forget.” The 128K context window of GPT-4 and the 200K window offered by some Claude models are, in essence, attempts to expand this form of short-term memory.
Humans work around limited capacity through chunking. AI systems use techniques such as retrieval-augmented generation, or RAG, and external memory to work around context limits.
Layer 3: Understanding — Working Memory
Next, you begin thinking about what the author means. You connect the sentence in front of you to earlier passages, general knowledge, and your own experience.
This is working memory. An executive network involving the prefrontal and parietal cortices holds information in mind, manipulates it, compares it with existing knowledge, and reasons about it.
Psychologist Alan Baddeley argued that working memory is not simply a temporary storage box. It is a coordinated system: a “central executive” directs attention while specialized subsystems handle verbal and visuospatial information.
How AI copies it
In AI, step-by-step reasoning, task planning, and the integration of multiple information sources can be viewed as engineering analogues of working memory.
Reasoning frameworks such as chain-of-thought and tree-of-thought aim to reproduce its active processing function. The Transformer’s famous attention mechanism offers a mathematical way to determine which parts of the input should influence processing most strongly. Memory-management frameworks such as Mem0, Zep, and Letta help AI preserve relevant context across longer interactions.
At this point, we move into the second stage: forming long-term memories.
Layer 4: Remembering — Episodic Memory
You reach a powerful scene in the book, and your hippocampus begins binding the experience into an event tagged with a particular time and place.
Years later, you may still remember sitting by the window while rain fell outside.
This kind of autobiographical, context-rich memory is called episodic memory.
It is the layer H.M. lost. Without a functioning hippocampal memory system, new events could no longer be properly encoded and consolidated. He could no longer remember what had happened to him.
One detail is especially important: the hippocampus does not simply store a perfect copy of an experience. It helps bind and index elements distributed across the brain. It works more like a library’s card catalog than the library itself.
The sights, sounds, and feelings of an experience are represented across different systems. Later, a cue can help the brain reconstruct the event.
How AI copies it
This is exactly the problem many AI memory systems are trying to solve.
A system might extract facts from a conversation — “the user is building a memory product” or “the user prefers concise writing” — attach information about time, source, and relationships, then retrieve those facts when a later task makes them relevant.
RAG systems follow a strikingly similar pattern: they divide documents into chunks, create embeddings, extract entities, and build relationships. The result is a searchable index of information.
Claude’s and ChatGPT’s memory features also aim to help AI recall useful information from previous interactions. Vector indexes and knowledge graphs function, in effect, as an engineered hippocampus.
Layer 5: Knowing — Semantic Memory
A month later, you may no longer remember which afternoon you read the book or which coffee shop you were sitting in. But its central ideas have become part of what you know.
This is memory consolidation. Over time, a memory can lose much of its original context and become generalized knowledge. Knowing what something is no longer requires remembering when you learned it.
How AI copies it
This is AI’s long-term knowledge layer. Vector databases such as Pinecone, Chroma, and Weaviate store semantic embeddings. Knowledge graphs such as Neo4j store relationships among concepts.
The architectural progression from temporary retrieval to persistent knowledge echoes, at a high level, the brain’s movement from context-bound experiences toward more generalized knowledge.
This is also where long-term AI memory becomes difficult. A database may hold billions of records, but storing something is not the same as remembering it well.
If a system cannot determine which information is trustworthy, important, and still current, adding more memories may only create more noise at retrieval time.
That requires a sixth capability: emotional and value-based weighting.
Layer 6: Caring — Emotional Memory
You read a passage and suddenly feel tears in your eyes. At that moment, the amygdala helps tag the memory with emotion: This moved me.
Emotional memory is less a completely separate memory system than a form of weighting applied to other memories. The amygdala and hippocampus work closely together: one helps represent what happened; the other helps signal what it meant to you.
The stronger the emotion, the more persistent the memory tends to be. That is why you may not remember what you ate for lunch last Tuesday, yet can recall the details of an embarrassing moment from ten years ago.
Emotion is the brain’s way of saying: This matters.
How AI copies it
Most AI systems still lack a mature version of this importance filter.
Companion products such as Character.AI and Replika come closest to giving AI an “amygdala.” They track aspects of a user’s emotional state and preferences, then adjust their responses accordingly.
This may be the most overlooked layer in today’s AI memory stack.
Many systems treat memories too equally. A crucial personal detail and a throwaway comment may receive nearly the same storage treatment. Yet weighting memories by emotional and personal significance could be essential to genuinely personalized AI.
Finally, we reach the third stage: turning memory into action.
Layer 7: Fluency — Procedural Memory
Think about reading again.
When you first learned, recognizing each word required effort. Years later, scanning a page, parsing sentences, and identifying key ideas happen almost automatically.
This is procedural memory: memory for how to do things.
The basal ganglia and cerebellum help turn repeated practice into skills that require less conscious effort. Once deeply learned, those skills can be remarkably durable.
That was the secret behind H.M.’s improving mirror-drawing performance. His surgery had not destroyed the brain systems that supported this kind of learning, so his skill continued to improve even without a conscious memory of practicing.
How AI copies it
AI agents are undergoing a similar transition.
At first, we wrote long prompts that explained every step: do this first, then do that. Over time, common operations were packaged into skills, while tools such as Zapier and n8n encoded repeatable processes as workflows. Systems began to learn which tools to call for a given class of task and which sequence of steps to follow.
AI is no longer limited to “knowing information.” It is beginning to accumulate reusable ways of getting things done.
The progression from manually explaining every step in a prompt to automatically triggering a packaged skill mirrors the human transition from effortful, conscious action to automatic performance.
Episodic memory allows AI to remember what we have experienced together. Semantic memory helps it know who you are and what you are working on. Procedural memory helps it learn how to work with you next time.
Only when these capabilities come together can AI evolve from a one-off question-and-answer tool into a genuinely long-term assistant.
The Seven Layers of Memory — and Their AI Counterparts
Notice something about the products in the final column: each one occupies only one or two cells.
When you read a moving passage, however, seeing (L1), understanding (L3), remembering (L4), learning (L5), and feeling (L6) can all happen at once.
The brain’s seven layers behave like a deeply integrated orchestra. Today’s AI stack still resembles a collection of soloists, each playing in isolation.
- It can preserve chat history without knowing which sentence mattered.
- It can retrieve old documents without recognizing that the information is outdated.
- It can remember a preference without developing a better way to work with you over time.
The next step in AI memory, then, may not be simply to store more. It may depend on solving four harder problems:
- Selection: What deserves to enter long-term memory?
- Consolidation: How should scattered experiences become stable knowledge?
- Updating: When facts and preferences change, how should old memories be revised?
- Forgetting: What should be deprioritized, allowed to expire, or permanently deleted?
Forgetting is not a bug in a memory system.
Often, it is what makes a good memory system possible.
Who Is Copying Whom?
At this point, you might ask: won’t AI memory eventually surpass the human brain?
In some respects, AI already has overwhelming advantages. A vector database can hold billions of records and retrieve a specific fragment in milliseconds. The same memory can be synchronized across countless instances. Incorrect information can be revised, deleted, or rewritten — while a human being cannot simply erase a traumatic memory.
But the human brain still holds several cards AI has not learned to play:
- Strategic forgetting. The brain suppresses or discards irrelevant information to make retrieval more efficient. For AI, storing everything can become a burden. If everything is remembered, nothing is remembered clearly. Forgetting is a feature, not a bug.
- Emotional weighting. The brain uses emotional intensity to help decide what may matter for a lifetime. AI has no mature equivalent.
- Deep integration. Perception, language, emotion, knowledge, and action continually shape one another. The layers do not operate as isolated modules.
- And, most fundamentally, imagination. AI generates from patterns learned during training and from the context it receives. The human mind can use memory as raw material for connections and ideas that feel radically new.
When people say that the 21st century will be the century of biology, the point is not necessarily that biology will replace AI. It may be that AI will increasingly learn from biology.
From the Atkinson–Shiffrin memory model to vector databases, and from H.M.’s role in revealing the function of the hippocampal system to the indexing logic of RAG, advances in AI memory repeatedly echo lessons already embodied in the brain.
The blueprint for the next step may already be in front of us: not another isolated breakthrough in one layer, but a fully coordinated seven-layer “digital brain.”
One Final Thought
Streaming media and the pace of modern life have trained us to live in fragments. We scroll through information while keeping several unfinished tasks in mind, constantly switching from one thing to another. We often describe this loosely as feeling “ADHD.”
Interestingly, the agents around us now suffer from something similar.
One moment, we are writing code in Codex. Then we switch to WorkBuddy to build a presentation. Next, we ask Claude how to plan the project’s next phase.
In the age of AI agents, work is naturally multithreaded and information is naturally fragmented. The problem is that these agents do not share a common memory.
Explain your background in one conversation, and you have to repeat it in the next. An agent may be capable of completing a task, yet its output still feels slightly off because it does not know what you did elsewhere, what you already tried, or which decisions you made.
Ultimately, the memory problem facing AI agents resembles our own: there is too much information, scattered across too many places, with no layer capable of connecting it all.
That is the gap we want to close.
Our team is currently building an AI memory product based on this seven-layer map. The goal is to make AI less like H.M. — requiring a fresh introduction every day — and more like a true second brain that remembers you and understands you better over time.
As you move between agents and tasks, it will provide the missing memory layer: preserving, retrieving, and reusing context that would otherwise remain scattered, so you no longer have to start from scratch in every conversation.
It will do more than archive chat logs. Over time, it will learn your projects, preferences, and ways of working, allowing each new interaction to build on everything that came before.
And instead of occupying a single cell in the table above, it will act as the connective layer across the entire system.
More soon. If you are interested in AI memory, follow me — I will continue sharing what we learn and how the product develops.
And if you are exploring this field yourself and want to build something meaningful, I would love to hear from you.
References
- Atkinson, R. C., & Shiffrin, R. M. (1968). Human Memory: A Proposed System and Its Control Processes.
- Scoville, W. B., & Milner, B. (1957). Loss of Recent Memory After Bilateral Hippocampal Lesions.
- Tulving, E. (1972). Episodic and Semantic Memory.
- Baddeley, A. D., & Hitch, G. J. (1974). Working Memory.
- Miller, G. A. (1956). The Magical Number Seven, Plus or Minus Two.
