Memory Is Not a Replay: Charan Ranganath on Forgetting, Imagination, and False Memories

Lex Fridman opens with Daniel Kahneman's split between the experiencing self and the remembering self: the satisfaction we take from a decision often comes from what we later remember, not from what we lived through in the moment. Charan Ranganath, a psychologist and neuroscientist at UC Davis, took Kahneman's class as an undergraduate at Berkeley, long before the Nobel Prize, and still calls it mind-blowing. The conversation they are having, he notes, could be retold in ten minutes, and that retelling would miss most of what actually occurred. Memory is not a replay of experience. It is biased toward beginnings, endings, peaks, and the best and worst stretches, and those are the pieces we consult when we choose.

That bias looks like a defect until you ask what the brain is for. Biologically the past is finished; what matters is extracting whatever helps make sense of the present and predict the future, cause-and-effect among them. We do everyday fortune telling before we even walk into a concert: the band will start late, there will probably be an encore, people will hold up phones. The valuable part of the illusion that we know what is coming is that, when it breaks, we get information. In information theory, Ranganath reminds Fridman, information is what you did not already have — so prediction error is where learning happens.

Asked how to live given that memory can be a primary source of happiness, Ranganath has been thinking more and more about a memorable life. During the pandemic, stuck inside looking at screens with the same people, he formed almost none of the good memories he means. His parents worked hard and rarely took vacations; now he takes his family to interesting places even when it is a pain, because the experiencing self will suffer and the remembering self will be glad. He likes to say there is no point in suffering unless you get a story out of it.

He illustrates with a near-death paddleboard outing with his friend Randy, a computational neuroscientist: pinned in a current, not a strong swimmer, watching Randy get pushed past him, thinking he might die alone. Everything that could go wrong almost did, and the story got better each time he told it in the lab. The photos of generators hanging over the water and metal gratings became a bonding artifact. The lesson he actually took away was that his ability to get something done in a crisis is better than he gives himself credit for. Fridman adds that even periods of depression can be motivating if you treat the darkness as material for the story on the other side.

Memory shapes bar choices and life choices, but Ranganath walks back a line from his own book. He had written that memory determines who we are; he now thinks the self is an evolving construct and that we underestimate its plasticity. He also thinks adolescence is far more important than the popular emphasis on the first three years. A great deal of mental illness begins to emerge then, and schizophrenia is now treated as a neurodevelopmental disorder because its window falls in adolescence and early adulthood. Parents feel every decision is carving the child, yet peers matter enormously, and much of what school actually teaches is how to get along and who you are — which can be traumatizing even with perfect parents.

Why remember almost nothing from the first years? He had just been talking with colleague Simona Ghetti. Infantile amnesia covers the first two years, with essentially no episodic memories; childhood amnesia covers the next two, which are fragmentary. One reason is that the hippocampus, crucial for remembering events, takes time to develop. Another is that the neocortex is so plastic, and a child's knowledge so rapidly rebuilt, that recovering the first patterns is like training a network on a couple of samples and then bombarding it with years of data: the roads back have been completely rerouted. A third is that a child's sense of self takes a while, so early learning is more "what happened" than "I remember, I was there."

Fridman offers a romantic guess: maybe the prediction-error rate in infancy is so high that the brain seals that chaos away as trauma. Ranganath does not like the word. Everything is changing — a baby spends the first three months discovering its toes — but children meet that waterfall with extreme curiosity and openness. Prediction error, for them, is not pain; it is the information they are seeking. The prefrontal cortex, which locks attention onto long-term goals, is still being tweaked into late adolescence and early adulthood, then has perhaps a decade of prime function before it declines.

That short "optimal" window for episodic memory used to bother him at talks. Then he redefined optimal. For most of human history, young adults with children needed a strong prefrontal cortex to hunt, forage, and hold the long haul; children in a relatively safe village needed to explore and play without being constrained by goals they do not yet have. Later in life, drawing on Alison Gopnik, he argues the job is no longer forming new episodes but passing on semantic memory — knowledge of the world, culture — to the next generation, a role still respected in indigenous cultures. Orca pods, another species with menopause, are led by grandmothers who transmit hunting and play traditions. Memory, he says, is doing what it is supposed to at every stage.

Asked how the brain forgets, Ranganath starts simple: memories are not photos on a phone but overlapping traces in a distributed pool of neurons, so competition can wipe things out. Sometimes the biological signals that promote long-term retention never arrive. And often we simply cannot find the cue that would send us back. In a neural network there is no file labeled "this memory"; the whole ecosystem lives in the weights, and what you extract depends on how you feed it. Fridman translates: you need the right query, the right prompt — which is also why a smell or a song can yank someone back a decade.

Memory researchers love to carve the system into pieces, but in the real world they connect. Working memory, a term coined by Alan Baddeley, is the ability to keep information online and control its flow; Baddeley split a passive store from a central executive identified with prefrontal cortex. Ranganath's early work argued that this short-term system is not independent of long-term memory: executive function itself requires learning and synaptic change. Semantic memory is general knowledge; episodic memory is memory for specific events. The internal model you load for a birthday party — eyes already going where the cake will be — is a working-memory product built from both, and it heavily determines what later gets called long-term memory.

He used to think the brain encoded episodes constantly, which would be enormously redundant. Following Jeff Zacks and collaborators including Ken Norman and Lila Davachi, his lab now argues that internal models update at points of high prediction error, uncertainty, surprise, or motivational significance — and those are the moments worth storing. The hippocampus and related networks show a bump of correlated activity at event boundaries, such as crossing from one room into another. If the model is good throughout an event you stay in predictive mode; at the boundary you encode, retrieve, and ask what is going on.

Can memory be trained? It depends, Ranganath says, on what you mean by optimal. In Why We Remember he argues you do not want to remember more, you want to remember better: focus on what is important, which is what brains are designed to do. Hermann Ebbinghaus, in the earliest quantitative studies, tried to memorize arbitrary nonsense and lost about 60 percent within a day, using a generous measure; nobody has really violated that basic fact. If your target is remembering everything, you have already set the wrong objective. The design principle is less is more: estimates he has seen put the human brain at something like 12 to 20 watts.

On that budget the only strategy is reuse. Neuromodulators such as norepinephrine, dopamine, and serotonin are released in biologically significant moments of surprise, fear, stress, and motivation, promoting lasting plasticity — a chemical stamp that says prioritize this. Attention is the other lever: Amishi Jha trains it with mindfulness in groups including Navy SEALs, and Adam Gazzaley uses video games, though transfer beyond the game remains controversial. Expertise, though, clearly trains attention. In a Ranganath lab study, as people became experts at identifying made-up objects, the biggest activity increases were in prefrontal cortex, not visual cortex, matching chess research and professional athletes who look three steps ahead.

Memory athletes were not born with extraordinary memory; they practice strategies until they can deploy them instantly. Yanjaa Wintersoul memorized an entire IKEA catalog in a viral video. A strategy for a deck of cards will not automatically help you navigate Austin, but whatever you care about, you can optimize, and that is a natural byproduct of expertise. Fridman uses Anki for spaced repetition and has played with the memory palace. Ranganath puts both in the same bucket: they work because they organize, not because they are magic.

Fridman is terrible at names: social anxiety eats the introduction, then the name is gone. Ranganath has the same blackout. The reason, he says, is that remembering a name is not remembering a name; it is binding an arbitrary label to a face and an identity, and brains are bad at meaningless pairings. Centuries ago a Miller might have been grinding flour; most modern names are pure symbols with nothing to latch onto. The fix is to build a connection somehow. He invents one on the spot: Lex Luthor, bald supervillain — Fridman is not bald, but the weird association thickens the network.

The other half of the trick is a preexisting hook. A generic John Smith becomes memorable at a conference if Ranganath has already read John Smith's papers. That is also why the method of loci, or memory palace, works: you already have a strong spatial structure, often a childhood home, so you can drop arbitrary items into locations and walk the path to retrieve them. Fridman notes a limitation: there is a sequence; you cannot just jump upstairs. Ranganath is unbothered. Some theorists think episodic memory evolved from spatial navigation; whether or not that evolutionary story is true, the method is not special. If you cannot visualize, stories work, and he can still sing an obscure Albania song from an episode of Cheers.

He uses a desk of sticky notes as a metaphor. If the desk is covered in yellow Post-its and your bank password is on one of them, good luck; if that note is hot pink, it pops — that is distinctiveness, which is especially useful for name-face pairs. The other method is organization: finances in one pile, to-dos in another, so you know where to look. Memory palaces supply a prestored structure of the second kind. Competing memories bury one another; unique processing and clean organization are how you find the note you need.

If Ranganath has an hour to memorize Spanish words, two half-hours a week apart beat two half-hours the same day for later retention. The boundary condition is when you need the information: cramming in the hallway right before a test puts the material right there, which is why he cannot remember much from college. Spacing helps later; massing helps now. In a computational model his group has under revision, this conversation is bound to a context — guitar sculptures, a big light umbrella. Later, at home with a beer, the current context does not match, so there is error, and the model erases or weakens the parts tied to a specific place and time while heightening the content.

Recall the same event in many times and places, and it becomes accessible in many times and places because it is not overfitted to one context. That is also why the memories we call on most start to feel like things we read about: you do not vividly re-experience them; they arrive as facts. The opposite is the smell, the song, the place you have not visited in years — boom, the episode returns, still glued to its original context. Fridman draws the implication: spaced repetition's power is that you lose attachment to a particular scene, and with it the intensity of the flavor. Anki is well suited to languages and semantic knowledge, not to keeping an afternoon intact.

A related model in PLOS Computational Biology addresses the testing effect: testing yourself on Spanish words beats restudying them. Traditional learning theory finds this odd — why add extra error instead of giving yourself a perfect replica? Ranganath's answer is that mismatch exposes bad connections, conceptually similar to backpropagation: keep the parts of the cell assembly that serve the memory and drop the rest. If you do not stress-test, the error is never fully exposed. His practical rule is that if you feel as if you are coasting, you are not learning, and the demoralizing "I thought I knew that" is the signal. You do not star in a movie by sitting around reading the script.

Fridman recounts telling Joe Rogan that Diego Maradona was still alive, a year after tweeting his grief; in his mind he had gone back to the runs on goal, and the man was living. Ranganath points to Marcia Johnson, a mentor he calls one of the greatest memory researchers, who asked how we tell imagination from memory when both are largely constructions. The prefrontal cortex can check sensory detail and place and time, and imagined events usually lack that vividness — but the check is slow, effortful, and imperfect. Imagination is the opposite move: take information from memory, recombine it, and throw it out. Patients with hippocampal amnesia struggle to imagine vivid future scenarios; they are stuck in the present.

Frederic Bartlett rejected quantifying memory, came from anthropology, and had people recall stories and poems. What they produced did not reflect all the details; it was filtered through prior knowledge, culture, and belief. He called remembering an imaginative construction: we do not replay the past, we imagine how it could have been from bits that come up. When people remember past events in the scanner, the default mode network — often dismissed as a daydreaming circuit — lights up and interacts closely with the hippocampus. Maps of people imagining possible or even implausible futures look, to the naked eye, almost the same. Ranganath's theory is that the network takes experiences apart into Lego blocks you can reassemble into the past or into an event that has not happened yet.

The Russian mnemonist Solomon Shereshevsky, described by Luria, had synesthesia that produced wildly vivid associations he used to memorize, and by some reports it haunted him: he had trouble telling imagination from reality. Ranganath notes that psychosis involves learning connections from prediction errors you probably should not learn, and confusing internal signals with the outside world. Genius and psychosis may be two sides of a double-edged sword. Memory athletes, meanwhile, usually buy a book, practice arbitrary strings of numbers and cards, and are not born special; Joshua Foer's Moonwalking with Einstein and Scott Hagwood, who learned the skills to fight chemo brain, are the pattern.

Ranganath himself has moved toward naturalistic tasks, because memory in the wild is a mix of interpretation, imagination, and perception. Once you know one IKEA, you can find the wine glasses in another without building a brand-new map. He suspects cognitive maps are economical graphs of landmarks and inferences, not Euclidean photographs. Hippocampal sharp-wave ripples — also seen in deep slow-wave sleep — look like a compressed play of earlier sequences and open a window of communication with neocortex. One role of sleep, in his theory, is connecting different events so you can build an internal model of how a person behaves after you have seen them now, next week, and a month later.

Ranganath got into the field because fMRI was taking off in graduate school: a whole human brain, no electrodes, people watching movies or memorizing words. A person lies in a 3 Tesla magnet; water molecules and the magnetic difference between oxygenated and deoxygenated hemoglobin become a tracer. When local activity rises, blood flows in, deoxygenated hemoglobin drops, and signal goes up. The maps look unimpressive — tiny intensity changes of about 1 percent — but those changes can predict whether someone will remember later. The coolest move is decoding: patterns of bright and dark voxels become, in his phrase, a QR code for a memory, so the house you grew up in and the car you drove at sixteen have different little barcodes in hippocampus.

Think of a computer folder you can sort alphabetically or by date; each brain network has a sorting scheme. In a favorite study by former postdoc Zach Raya, movies showed two people in two cafes and two supermarkets: one network reused a pattern whenever Alex appeared, wherever he was; another reused a pattern for a given supermarket, and the same patterns showed up during recall. Those are the building blocks; hippocampus only has to supply the blueprint for putting them together. fMRI is slow — the hemodynamic response peaks about six seconds after a burst of activity — and more sensitive to excitability and glial metabolism than to spikes. For stories and podcasts, a lot of the action is in that slow integrative timescale.

Asked for the flavor of discoveries in the science of memory, he lists forgetting curves, the power of schemas, expertise and its distortions, and the fact that the act of remembering can change a memory, strengthen or weaken neighbors, and turn memory into an ecosystem. Breaking continuous experience into discrete events was another. The kitchen example finally arrives: you walk in, forget why, grab food from the fridge, sit down, and remember you went for your watch. Crossing a doorway updates the internal model, so you have to mentally time-travel to the previous room. Hippocampal activity at those event boundaries, while people merely watch a movie, later correlates with their ability to remember a story outside the scanner.

Mice let you turn on a particular memory by activating the neurons that were active during an experience, and some learning genetics are old. But mice explore by smell, their vision is built to catch death from above, they lack language and human social structures, and a lab mouse's world is impoverished next to an adult human. Timescales differ: primate brains hold information across longer stretches. Ranganath is blunt about colleagues who think you can learn everything about memory from a mouse. You go to mice for how neurons talk on a millisecond scale; you still need humans for the rest.

Surveys suggest about 75 percent of people have had déjà vu; Ranganath polls his classes and most raise their hands. Hughlings Jackson noticed intense déjà vu as an aura in some epilepsy patients; Wilder Penfield later elicited the feeling, and sometimes a full re-experienced memory, by electrically stimulating temporal lobe. Ranganath likes a fluency account: higher-order areas tune themselves to familiar input, so you do not scan a known room with your eyes, and that ease feels like "I have been here." Anne Cleary at Colorado State ran people through a virtual museum, then a virtual arcade with the same map under a different skin — carpet, color, theme all changed. Most had no conscious idea the layouts matched, yet reported intense déjà vu: a partial match that solicits familiarity without recollection.

On false memory he likes to say there is no such thing as true or false memories, quoting Johnny Rotten: he does not believe in false memories any more than he believes in false songs. A memory mixes bits of what happened with inferences and theories, the way a scientist mixes data with a theory. Small distortions fill gaps: a story in which a doctor says things are "very much like you would have expected" gets remembered as the doctor telling the patient he had cancer. Things get out of hand when people cannot tell imagination from occurrence.

Elizabeth Loftus pioneered the misinformation path: you remember an event, someone tells you something about it, and if you cannot source the extra information it mixes into the original story. Repeat the process and the memory detaches from what happened. In the lab, a chunk of people can be led to stitch together a vivid memory of something that never took place, because they are remembering the act of trying to remember. Fridman notes that propaganda aims to do this at scale, and Ranganath agrees: every foible of human memory is magnified in social interaction.

Social contagion is misinformation spreading like a virus: once two people share a memory it is no longer yours or mine. Viet Thanh Nguyen's family called the conflict the American War, not the Vietnam War — a valid perspective that stunned Ranganath — and dominant narrators make groups remember less, while diverse voices yield more accurate memory if people will tolerate discomfort. His fear about AI is not what models will do but what people will do: email autocomplete was exactly right a third of the time, wrong a third, and "good enough" a third, quietly constraining language, which is why his mantra is to diversify your training data. Life may be better than fifty years ago without people being happier, because we learn mainly when rewards deviate from expectation, and dopamine looks less like pleasure than like the discomfort that energizes seeking. Matthias Gruber found that curiosity itself — the question, before any answer — lights up dopamine-related circuits, a gap you can use as fuel or as an excuse to return to the echo chamber.

Fridman turns to 1984: can torture make people remember anything? Ranganath points to "the third degree" — beatings, starvation, physical demands — and to research on enhanced interrogation, which tends to produce what the interrogator wants to hear. The recipe is specific: an authority figure says you did this and we have witnesses, so you start to doubt yourself; stress from hunger, cold, or unbearable music downregulates prefrontal cortex; then they get nice, walk you through how it could have happened, and feed you the information. In that weakened state, encouraged to imagine a plausible script, some people can be coaxed into a memory of something that never occurred. There are confessions later overturned by objective evidence; he mentions a sheriff who came to believe he had a memory of sexual abuse.

Why do we remember heartbreak? Attachment is biologically significant, so memory is designed to capture it, especially when beliefs about someone collapse or regret turns into rumination. His first cat, a wedding present, died of FIP at four; he saw her everywhere in the house, and later a dog that died of cancer still appears. From clinical training he takes the view that much of therapy is reframing: not rewriting memory or imposing an interpretation, but offering a perspective more oriented toward learning and gratitude. Fridman cites Louis C.K.'s show, in which an older man tells the heartbroken lead that heartbreak is the best part, because you intensely feel how valuable the love was; the worst part is forgetting. Ranganath asks whether the owner of a broken heart is better than the owner of a lonely one; Fridman says yes, then jokes he will need more Bruce Springsteen to be sure.

After turning 50, Ranganath thinks about death often: fewer years ahead than behind, and which memories he wants to carry. A Buddhist he spoke with said the way to renounce attachment to loved ones is to appreciate this moment, knowing they will die. Fridman meditates on mortality daily but is torn about removing himself from the world's emotions. Ranganath contrasts strains of Hinduism via the Bhagavad Gita: do not renounce action, renounce the fruits of action; do the process because it is part of the world's balance. He tries to let go of whether the book sells or whether he made Fridman laugh, and just share information — hard, because we are so driven by the outcome.

Studying memory warps time: one hour ago and two hours ago feel very different, while one year ago and one year plus an hour do not. During remote teaching in the pandemic he polled his class: almost everyone said the days were slower, and a majority said the weeks were faster, which is nonsense in physics and natural in memory — same context, same activities, one long event, and at week's end nothing to retrieve. Nostalgia, Felipe De Brigard reminded him, was coined by a Swiss physician studying traumatized soldiers as a disease; in older people it can lift mood, but it can also become "those were the good old days and they are over." He lived through the 1970s: gas lines, fear of Russia and nuclear war. Selective remembering is useful if it brings present happiness and toxic, personally or collectively, if it narrows the present without your noticing.

On brain-computer interfaces and Neuralink, Ranganath has not studied the company closely: interesting surgical robots, but as a scientist he is not seeing published evidence they are near the academic state of the art. The field is much bigger than one firm. Colleague Sergey Stavisky is working on speech prosthetics that combine decoding with AI, and on movement, at an enormous rate of progress. The deeper issue Stavisky raised is ethical: language includes intentions and internal models, and reading those out is a different act from decoding sounds and movements. Fridman pushes the extreme: if neuroscience can map QR codes for political thoughts, McCarthyism gets a new instrument, and freedom of thought becomes tricky once thoughts can be collected.

Can we turn an unpleasant memory pleasant? Behaviorally, yes: under some conditions misinformation can swap people and places. A harder tool is reconsolidation — when you recall a memory, connections in that cell assembly become more modifiable, and some groups try to reduce the visceral component of fear memories in that window. As an outsider looking at the data, Ranganath finds mixed results, partly because you must fully recall the traumatic memory before you can modify it, and "what is the memory?" is the real problem. Watching the same movie, two people can recall the same thing in different words and durations; the thought is deeper than the words. Crude behaviorism — give A, get B — is, in his view, a bankrupt concept of memory.

A discussion between neuroscientists and AI people turned on why city driving is harder than highway driving: the car can do the driving, but it cannot predict pedestrian intentions. Fridman has watched intersections for hours and concluded that every one has a personality — a history of jaywalking, regulars getting off the subway, newcomers following whoever goes first — different in Boston, New York, Austin, San Francisco. He thinks enough data compressed into a representation is understanding, the way an LLM generating a trip to the kitchen has to have a world model that includes cruelty and wonder. Ranganath's vivid observation: once you are a good driver you can see a mile away which car is going to pull an aggressive move, a glowing icon in peripheral vision while you ignore the rest. Attention on a budget solves that puzzle with a small allocation of resource.

OpenAI had just announced memory. How hard is it to put human memory into AI? Superficially not that hard; at a deeper level, very hard, because we do not understand episodic memory. Steve Grossberg's stability-plasticity dilemma asks when you overwrite and when you make incremental changes: English has too many exceptions, and learning them fast can cost you the rules. David Marr, Jay McClelland, Randy O'Reilly, Neal Cohen and others proposed two systems — a fairly dumb episodic store that says this happened once, and accumulated semantic knowledge — so a violation of expectation can be stored and used next time without erasing what you already know. The unsolved questions are when to record, how much, what to prioritize, when to use the episode, and when to throw it away, and he frankly does not know whether he wants a self-motivated LLM. Sam Gershman quipped that if you trained a typical AI to make you as much money as possible, the first thing it might do is sell your house.

Postdoc Tyler Bonnen took state-of-the-art computer vision models and gave them the object tests used with patients who have damage in memory-related areas. The images the models could not crack — more parameters and layers did not help — were the same ones that tripped people with perirhinal cortex damage, who could still succeed if given enough time. People never see every pixel at once: they move their eyes, grab high-resolution points, and direct the exploration, so integrating less information at each step across time produced something the models could not do. Fridman ties it to Transformers: attention is the mechanism, and how you allocate it is close to a definition of intelligence. Ranganath will not be convinced a simulator has replicated human minds unless it makes the same characteristic mistakes; he has not seen that in ChatGPT, and he does not even like the word intelligence.

ADHD is personal: 1970s school labeled him, a diet without artificial colors followed, and he still recognizes time blindness and impulsivity, with prefrontal balance between focus and shifting knocked off. Memory looks poor because attention is; interesting things stick. Science enables the profile — novelty, curiosity, nobody dies if you are late — but email still kills him, producing fragmentary memories of everything, so he blocks time for mail, writing, and goofing off as an investment. Fridman cites Cal Newport on task switching; Ranganath adds that conference laptops create artificial event boundaries, you run a few seconds behind, and work by Melina Uncapher and Anthony Wagner shows the memory damage is worse than people think.

Music is where he learned to love prediction error. Trumpet via sheet music, then guitar he could not get until Sonic Youth showed him you do not have to write music the way people tell you it should sound. He messed with notes that should not go together and with 5/4 and 7/4; Pink Floyd's "Money" feels like 4/4 but the riff is seven beats, a little error every cycle. He plays in Pavlov's Dogz, a cover band of memory researchers including MIT's Earl Miller on bass, and likes to collapse "I Wanna Be Your Dog" into feedback. Some people flee that dissonance; he gravitates toward it.

What he loves most about the mind is the internal model: everything you see, hear, touch, and taste, held together by a kind of dark energy without which experience is just a bunch of stuff. We sample a little of the world and extract a lot of meaning, imperfectly but well. He wrote Why We Remember in part so the public would hear from scientists, including the uncertainty — COVID was an example of scientists not being heard. Thomas Kuhn's The Structure of Scientific Revolutions is essentially his only philosophy: assumptions are the core of science, and finding a whole zone of what you do not know is the exciting part. Fridman leaves listeners with Haruki Murakami: most things are forgotten over time, but some memories we can never rub away, and they remain like a touchstone.