To continue providing free, value-first guides and curated resources, some of the links on this site are affiliate links. If you click through and make a purchase, we may earn a small commission at absolutely no extra cost to you, which helps support the platform.
AI and Memory: How to Use It Without Losing Yours
You can produce thousands of words with AI and remember almost none of them. Or you can use the same tools and come away sharper. The difference is not the AI. It is how memory actually works — and whether you are working with it or against it.
- The 83% problem
- How memory actually works
- Why AI bypasses the encoding process
- The hippocampus: your memory engine
- Why writing is a memory act, not just a communication act
- The dementia connection: what chronic underuse means
- What AI use that protects memory looks like
- Seven memory-protective habits for AI users
- Conclusion: memory is not a passive record
The 83% Problem
Imagine spending an hour producing a detailed, well-argued piece of writing — researching it, shaping it, reviewing it — and then being unable to accurately recall a single paragraph of it fifteen minutes later.
This is not a hypothetical. It is what an MIT study found happening to the majority of people who write using AI tools.
In the study, participants were divided into two groups. One group wrote essays using ChatGPT. The other wrote without any AI assistance. Both groups produced finished essays. Both spent comparable time on the task. And then researchers measured something simple: could the participants remember what they had written?
Eighty-three percent. Not a fringe result. The overwhelming majority of people who produced content using AI walked away having retained essentially none of it. They had the essay. They did not have the ideas.
This result surprises people who assume that reading something — even AI-generated content you commissioned — is sufficient for memory. It is not. And understanding why it is not requires understanding how memory actually works at the neurological level. Once you do, the MIT finding stops being surprising. It becomes the only result that makes sense.
How Memory Actually Works
Memory is not a recording. This is the single most important thing to understand about it, and it is the thing that most people’s intuitions get wrong.
When you experience something — read a book, have a conversation, solve a problem, write an essay — you do not record it the way a camera records a scene. There is no playback. Instead, memory is a reconstruction process: the brain encodes selected features of an experience into a distributed pattern of neural connections, and later “remembers” by reactivating and reconstructing that pattern. What gets encoded, how strongly, and for how long depends almost entirely on the quality and depth of the original processing.
Memory formation happens in stages, and each stage requires active cognitive participation.
The key word across all four stages is active. Memory is not something that happens to you when you are exposed to information. It is something your brain does with information it has been made to genuinely process. The depth of processing determines the depth of encoding, which determines how much you remember and for how long.
Shallow processing — skimming, passive reading, absorbing information without connecting it to what you already know — produces shallow encoding, which produces weak or absent memories. Deep processing — generating, questioning, connecting, applying, explaining — produces deep encoding, which produces durable memories that are accessible and usable weeks, months, and years later.
Why AI Bypasses the Encoding Process
With that understanding of how memory works, the MIT study result is no longer puzzling. It is inevitable.
When you write an essay yourself — genuinely, in your own words, working through the argument — you are engaged in one of the most cognitively demanding encoding activities available. You are deciding what matters and what doesn’t. You are constructing logical connections between ideas. You are translating abstractions into specific language. You are making hundreds of small decisions — this word, not that one; this example, not that one; this structure, not that one — each of which requires active engagement with the material. Every one of those micro-decisions is a memory encoding event.
The struggle of writing is not an obstacle to understanding. It is the mechanism by which understanding is formed and encoded.
When you ask AI to write the essay, none of that happens. You provide a prompt. You receive output. You read the output — at roughly the depth of attention you give to anything you read once, without particular purpose. The hundreds of encoding micro-events that the writing process would have produced simply do not occur. The hippocampus receives no signal to encode. The long-term storage transfer never begins. The ideas pass through your working memory and leave no lasting impression.
This is not a failure of attention or effort on your part. It is the predictable neurological consequence of removing the cognitive work that produces encoding. You cannot encode a process you did not perform.
You provide a prompt. AI generates the text. You read it, perhaps lightly edit it, and use it. Your brain processes the output at reading-comprehension depth — shallow encoding. No decisions were made, no arguments were constructed, no words were chosen. The hippocampus has little to encode. Most of the content is gone within the hour.
You draft your own version first, however rough. Your brain constructs the argument, chooses the words, makes the decisions. You then use AI to challenge and extend. You write your own synthesis. The original thinking has been encoded through the act of production. The AI interaction deepens and refines what is already there.
The distinction between these two modes is not philosophical. It is the difference between remembering what you know and producing content you will never be able to reconstruct. For casual social posts, the difference may be trivial. For professional knowledge, strategic thinking, learning, or any domain where you need to actually know and use what you produce — it is the difference between developing expertise and performing it.
The Hippocampus: Your Memory Engine
The hippocampus is a small, seahorse-shaped structure deep in the medial temporal lobe, and it is the single most important structure in the formation of new declarative memories — the kind of memories that contain facts, concepts, events, and ideas. Without a functioning hippocampus, new long-term memories cannot form at all. The famous patient H.M., who had both hippocampi removed in 1953, could not form a single new long-term memory for the remaining 55 years of his life. Every person he met, every conversation he had, every experience he lived was gone within minutes.
Most people will never experience hippocampal damage of that severity. But the hippocampus exists on a continuum of health and function, and its performance is highly sensitive to the inputs it receives — or fails to receive.
The hippocampus thrives on challenge, novelty, and active cognitive engagement. It is specifically activated by the experience of working through something difficult, by making connections between disparate pieces of information, by the spatial and conceptual mapping involved in constructing a complex argument or solving a multi-step problem. These are precisely the activities that deep engagement with AI — and the absence of engagement with passive AI use — either provides or removes.
The brain imaging component of the MIT study showed significantly reduced activity in the hippocampal region among participants who used ChatGPT passively. This is not a measurement of how much they remembered after the fact. It is a measurement of how much encoding activity was occurring during the writing process. The memory failure came later because the encoding work was never done.
The hippocampus is also one of the first structures to show damage in Alzheimer’s disease — which is why memory loss is typically the earliest and most prominent symptom. And hippocampal health, research consistently shows, is directly related to the level of cognitive challenge it receives across a lifetime.
The London taxi driver research cited in Post 2 is the most famous demonstration of this: years of active navigational learning produced measurably larger and more active hippocampi. The inverse is equally documented: chronic cognitive understimulation is associated with hippocampal volume reduction and increased dementia risk. Passive AI use, by reducing the hippocampal activation that active cognitive work provides, sits within this broader pattern of understimulation.
This is the link between the MIT study’s acute finding — people couldn’t remember what they just wrote — and the longer-term concern about brain health. The immediate memory failure is a symptom. The underlying cause is the chronic removal of the hippocampal challenge that memory formation and brain maintenance both require.
Why Writing Is a Memory Act, Not Just a Communication Act
Most people think of writing as a way to communicate ideas they already have. This is a partial truth that misses something more important. Writing is also — perhaps primarily — one of the most powerful memory encoding and knowledge-building processes available to the human brain.
The cognitive scientist David Geary distinguishes between what he calls biologically primary knowledge — the things the brain acquires automatically and effortlessly, like spoken language, basic social cognition, and emotional recognition — and biologically secondary knowledge, the kind that requires deliberate, culturally transmitted effort to acquire. Reading. Writing. Mathematics. Scientific reasoning. These are not natural to the brain. They are achievements of civilization, and they require effortful practice precisely because the brain was not evolved to do them automatically.
Writing is among the most demanding of the biologically secondary skills. Constructing a coherent written argument requires sustained attention, working memory, planning, revision, and the kind of metacognitive monitoring — am I saying what I mean? does this follow from that? is this the right word? — that exercises the prefrontal cortex, hippocampus, and language networks simultaneously. It is, neurologically, a full-system workout.
This is why many of the most productive thinkers in history were also prolific writers — not because they had more to say, but because the writing was itself the engine of their thinking. Darwin wrote constantly in notebooks. Einstein wrote thought experiments in longhand. Toni Morrison described writing not as communicating what she knew but as discovering what she knew. The act of writing was not downstream of the thought. It was the process by which the thought became clear.
When AI writes for you, this process is interrupted at its source. The thinking that writing produces — the clarification, the discovery, the encoding — does not happen. You receive the output of a process you did not perform and therefore cannot own in any meaningful cognitive sense. It is the equivalent of hiring someone to do your exercise for you and expecting to get fit.
The specific encoding mechanisms that writing activates
Generation effect. Information that you generate yourself — rather than simply receive — is remembered significantly better. This has been documented in hundreds of studies since the 1970s. The act of producing a word, a phrase, or an idea creates a stronger memory trace than the act of reading the same word, phrase, or idea. This is why active recall outperforms re-reading as a study method, and why writing your own version of something encodes it far more deeply than reading someone else’s.
Elaborative encoding. When you write, you are constantly connecting new ideas to things you already know — choosing examples, drawing analogies, anticipating objections, providing context. Each of these connections is an elaborative encoding event. The richer the connections, the more durable the memory. This elaboration is entirely absent when AI generates the connections for you.
Desirable difficulty. Cognitive psychology has established that the difficulty of a learning or encoding task — up to a point — improves long-term retention. Struggling to find the right word, to construct the right argument, to organize ideas coherently is not a sign that learning is going badly. It is the sign that encoding is happening deeply. Making the task easier — as AI does — reduces the desirable difficulty and weakens the resulting memory trace.
The Dementia Connection: What Chronic Underuse Means
The acute finding of the MIT study — poor memory of just-completed AI-assisted work — is concerning enough on its own. But the longer-term implications, connecting habitual passive AI use to dementia risk, require us to zoom out to a different timescale.
Dementia is not a sudden event. It is the endpoint of a process that unfolds over decades, shaped by the accumulation of neurological changes and the presence or absence of cognitive reserve. As we covered in Post 2, cognitive reserve is the brain’s resilience against age-related damage — the buffering capacity built through a lifetime of cognitive engagement that allows the brain to sustain function even as biological changes occur.
The activities that build the strongest cognitive reserve are, consistently, those that involve sustained, effortful, complex cognitive processing: education, cognitively demanding work, bilingualism, musical training, complex social engagement, and — critically — activities like writing, reading deeply, and active problem-solving. These are exactly the activities that habitual passive AI use systematically replaces.
No single AI interaction causes dementia. This cannot be said clearly enough. The risk is not acute; it is cumulative. It is the pattern over years and decades that matters, not any individual session. But the pattern that chronic passive AI use establishes — consistent removal of effortful cognitive processing from daily life — is precisely the pattern that cognitive neuroscience identifies as a risk factor for accelerated cognitive aging and reduced resilience against dementia.
Young adults who adopt passive AI use habits now, before the long-term consequences are visible, have the most at stake. The cognitive reserve that protects against dementia in your 70s and 80s is built substantially in your 20s, 30s, and 40s. The habits formed during those decades — of effortful or effortless cognitive engagement — shape the reserve you will either have or lack when you need it most. The window matters. It is open now.
What AI Use That Protects Memory Looks Like
Everything in this article describes what passive AI use costs. It is equally important to describe what AI use that protects and strengthens memory actually looks like in practice — because this is not a call to avoid AI. It is a call to use it in a way that works with your memory biology rather than against it.
Memory-protective AI use has three core characteristics. It maintains the encoding process by ensuring your brain is doing the cognitive work that encoding requires. It uses AI to deepen and extend that work rather than replace it. And it includes deliberate retrieval practice to consolidate what has been learned.
Draft first, always
Before consulting AI on any substantive task, spend time — even ten or fifteen minutes — producing your own rough version. This can be notes, a mind map, a rough draft, or simply writing out what you currently think and where you are uncertain. This step activates the encoding process. It creates a cognitive scaffold onto which AI-generated information can be attached and remembered. It is not about being right before you consult the tool. It is about giving your hippocampus something to work with.
Engage, don’t extract
When you receive AI output, resist the impulse to immediately accept and use it. Ask follow-up questions. Push back on claims you are unsure about. Ask it to explain its reasoning. Ask for the strongest objection to the argument it just made. This active engagement turns AI interaction from a passive reception event into an encoding-rich dialogue. The cognitive friction you generate by questioning and pushing back is the friction that creates durable memory.
Close and synthesize
After any substantive AI interaction, close the window and write — in your own words, without reference to the output — what you actually understood and what you now think. This single habit is the most important memory-protective practice available to AI users. It forces retrieval (strengthening the memory), requires translation into your own language (deepening encoding), and surfaces gaps in understanding (allowing targeted follow-up). It takes five to ten minutes. It is the difference between a conversation you will remember and one you will not.
After any AI-assisted work session, close everything and set a two-minute timer. Write down, from memory, the three most important things you just learned or decided. If you cannot do this, the session produced content without producing understanding. The test is not a grade — it is a diagnostic. It tells you whether the encoding process happened. If it didn’t, the synthesis step above is what fixes it.
Seven Memory-Protective Habits for AI Users
These habits are grounded directly in the memory science covered in this article. Each one targets a specific mechanism — encoding, consolidation, retrieval, elaboration — to ensure that your AI use builds understanding rather than bypassing it.
- Draft before you prompt Spend at least ten minutes producing your own rough thinking before opening AI. Notes, bullet points, a rough paragraph — anything that activates the encoding process before AI provides the output that would otherwise bypass it. This is the single highest-impact habit for memory protection.
- Explain it back before you use it Before applying or sharing anything AI produced, close the window and explain the key ideas in your own words, out loud or in writing. If you cannot explain it without looking, you do not yet understand it well enough to use it. Go back, engage more actively, and try again.
- Ask AI to quiz you After a substantive AI interaction on a topic you want to retain, ask the tool to test your understanding: “Ask me five questions about what we just covered.” Active retrieval is the most powerful memory consolidation tool available. Using AI to generate retrieval practice is one of the highest-value uses of the technology for learning.
- Use the teaching method After any learning session with AI, write a short explanation of the core ideas as if explaining them to someone who knows nothing about the topic. The process of simplifying and sequencing for an imagined audience forces elaborative encoding and surfaces the gaps in your understanding that passive reading hides.
- Review after 24 hours Memory consolidation happens primarily during sleep. Return to the key ideas from any important AI-assisted learning session the following day — not by re-reading the AI output, but by attempting to recall the key points from memory before checking. This spaced retrieval dramatically improves long-term retention compared to a single session.
- Keep a synthesis journal Maintain a simple running document — not a repository of AI output, but a record of your own thinking developed through AI interactions. What did you learn? What changed your mind? What do you now think that you didn’t think before? The act of writing this consistently is a powerful encoding, consolidation, and retrieval practice rolled into one.
- Preserve manual writing practice Identify at least one form of regular writing you commit to doing entirely in your own voice, without AI assistance. This might be a personal journal, a weekly reflection, a professional newsletter, or even long-form emails to people you respect. This is your hippocampal training. It is not optional maintenance. It is how you keep the most important cognitive infrastructure you have in working order.
Memory Is Not a Passive Record
The 83% of people who could not remember their AI-assisted essays are not unusual or inattentive. They are demonstrating what happens when information is received without being processed — which is what passive AI use, by design, produces. The brain does not record what passes through it. It encodes what it works with. And working with something requires exactly the kind of cognitive effort that passive AI use is optimized to eliminate.
This is not an argument against AI. Used actively — to challenge your thinking, to extend what you have already drafted, to deepen understanding you have already begun to form — AI can be one of the richest cognitive experiences available. It can expose you to perspectives, evidence, and arguments that deepen your understanding and, through active engagement, encode more deeply than solo reading or thinking would have achieved.
But that outcome requires the encoding process to be intact. It requires your brain to be doing the work of processing, connecting, generating, and questioning. It requires you to be an active participant in the cognitive experience rather than a passive recipient of its output.
Memory is not a passive record of what you have encountered. It is an active construction of what you have genuinely processed. Your relationship with AI determines, day by day and habit by habit, which of those you are building.
Post 4: The Politeness Principle — Why How You Talk to AI Changes Your Brain. The surprising neuroscience of why treating AI as a dialogue partner — rather than a search engine — produces better output, less cognitive fatigue, and a more brain-protective relationship with the technology.
0 Comments