Charan Ranganath on Human Memory, False Memories, and How the Brain Reconstructs the Past

Charan Ranganath, a psychologist and neuroscientist at UC Davis and author of Why We Remember, lays out how human memory works — not as a recording device but as a constantly reconstructed model of the past — and explores what that means for false memories, imagination, déjà vu, identity, and the limits of AI.

Key ideas

1. Memory reconstructs; it does not replay. The brain does not store and play back experiences like a video file. Every act of remembering is an active reconstruction, stitching together fragments of what happened with general knowledge, inference, and the demands of the present moment. The same memory, recalled under different conditions, will come back differently each time.
2. Episodic memory is encoded at event boundaries, not continuously. The hippocampus does not record a steady stream of experience. It marks moments of high prediction error or surprise — the points where the brain’s internal model breaks — and encodes those. Walking between rooms, encountering the unexpected, and crossing a narrative threshold all trigger encoding spikes.
3. The act of remembering changes the memory. Recalling an event opens its neural representation to modification. Introduce misinformation at the moment of recall, and that information can be woven into the original memory so tightly that it becomes indistinguishable from what actually happened. At scale, this is how propaganda rewrites collective memory.
4. Imagination and memory share the same neural substrate. Brain scans of people imagining future events and people recalling past ones look nearly identical. The default mode network — once dismissed as mere daydreaming — underpins both. Patients with hippocampal amnesia lose not only their past but their ability to construct a vivid future.
5. The brain is optimised to remember better, not more. Forgetting is not a bug; it is the brain’s economy at work. The goal is to retain what is causally relevant to the present and future, compress repeated experience into generalised knowledge, and mark genuinely novel events with durable episodic traces.

Content

The experiencing self and the remembering self

Ranganath opens with Daniel Kahneman‘s distinction between the experiencing self (what actually happens to you) and the remembering self (what you take away from it). A three-hour conversation can be summarised in ten minutes, but that summary misses most of the experience. Memory is not a record of experience — it is a highly biased sample, weighted toward beginnings, endings, peaks, and emotionally intense moments. This matters enormously for decisions: when we consult memory to choose where to eat, whom to trust, or how to feel about our lives, we are consulting a distorted digest, not a transcript.

The brain’s bias is not arbitrary. It is calibrated toward causal utility — the information from the past most likely to help us understand the present and predict the future. Routine events, by definition, teach us little. What stands out are prediction errors: the moments when what happened did not match what we expected. Those errors are the signal; the rest is noise.

How memory is built: hippocampus, event boundaries, and internal models

The hippocampus (a seahorse-shaped structure buried in the temporal lobe, central to forming new episodic memories) does not record experience in a continuous stream. Instead, it operates in a predictive mode: while events unfold according to expectation, the brain runs on its internal model with minimal new encoding. At event boundaries — a change of scene, a surprise, a shift in who is present — activity in the hippocampus spikes, and a new episodic memory is formed. Ranganath’s lab can predict, from hippocampal activity at these boundaries during a film, how well a person will remember its narrative later on.

This architecture means that a monotonous day feels slow hour to hour but leaves almost no memories at week’s end. Variety, novelty, and surprise are not just pleasant — they are the conditions under which the memory system works.

Memory also depends on the right cue at retrieval. The brain does not store memories as discrete files; they are distributed across overlapping networks of neurons. Two memories that share context can interfere with each other — the more crowded the space, the harder it is to find the one you want. The method of loci (the Memory Palace technique) works precisely because it imposes a pre-existing spatial structure onto arbitrary material, giving each item a unique cue that separates it from the others.

Imagination as recombined memory

Frederic Bartlett — a revolutionary memory researcher who rejected statistics and came from anthropology — showed in the early twentieth century that people’s recollections are not faithful transcripts but imaginative reconstructions filtered through prior knowledge and cultural expectation. Ranganath extends this: remembering and imagining are, at the neural level, the same process run in different modes. Both involve the default mode network disassembling past experiences into component parts — people, places, actions, emotions — and reassembling them according to a directive. For memory, the directive is ‘reconstruct what happened’; for imagination, it is ‘build something plausible or novel’. The Lego blocks are the same; the instructions differ.

The clinical evidence is stark: hippocampal amnesics cannot imagine vivid future scenarios. Asked what they might do after leaving the room, they produce flat, stereotyped answers — not because imagination is a separate faculty that has been damaged, but because memory and imagination share the same machinery.

Déjà vu

Déjà vu is a partial match. Higher-order areas of the temporal lobe integrate input from many sources, tuning themselves over time to process familiar configurations more fluently. When a novel scene contains enough structural similarity to a past experience — same spatial layout, same pattern of relationships — these areas generate a sense of familiarity without a retrievable specific memory. Ranganath cites Anne Cleary‘s virtual-reality experiments: participants moved through a virtual arcade that had the same layout as a museum they had visited earlier, with different surfaces and colours. They reported intense déjà vu but had no conscious recollection of the museum. The map matched; the skin did not.

Epilepsy provides a more extreme version: electrical activity in the temporal lobe can trigger déjà vu artificially, before or during a seizure. Early neurologists, including Wilder Penfield, elicited the experience by directly stimulating the temporal cortex. Déjà vu is not a mystical glitch — it is the familiarity signal firing without the recollection signal to complete it.

False memories and misinformation

Ranganath argues that the category ‘false memory’ is less stable than it appears. All memories are reconstructions blended with inference; the question is always how much the blend has drifted from the original event. Elizabeth Loftus pioneered the experimental study of this drift: expose someone to misinformation about an event they have already encoded, and that information can be incorporated into their memory of it so thoroughly that they cannot later separate what they originally experienced from what they were told.

The mechanism is compounded by reconsolidation: every time a memory is retrieved, the neural assembly encoding it becomes temporarily plastic. This is a vulnerability. Under conditions of stress (which suppresses the prefrontal cortex’s monitoring functions), combined with suggestion from an authority figure, people can be led to construct elaborate memories of events that never occurred. False confessions in criminal justice are one domain where this produces documented injustice. Propaganda is another: at the collective level, misinformation spreads through a population’s shared narrative like a contagion, each retelling incorporating a little more distortion, until the group’s memory is entirely detached from what happened.

The antidote Ranganath proposes is epistemic diversity: groups in which many voices are encouraged, and in which dominant narrators do not crowd out others, produce more accurate collective memories and more information overall from any shared experience.

Memory across the lifespan

The prefrontal cortex — critical for goal-directed attention, monitoring, and the executive control of memory — takes the longest of any brain region to develop, maturing into late adolescence and early adulthood. Children’s ‘openness’ and exploratory curiosity is not a deficit but an adaptation: a brain without a tightly set prefrontal cortex is free to learn without premature goal-commitment. Adolescence, conversely, is when the system is most vulnerable to disruption — schizophrenia, ADHD, and other mental-health conditions tend to emerge precisely during this period of reorganisation.

At the other end, older adults show reduced hippocampal activity at event boundaries and declining episodic memory. But Ranganath reframes this too, citing Alison Gopnik and research on orca pods: orca grandmothers, who undergo menopause and survive long past reproductive age, lead the pod and pass cultural traditions to younger generations. The elders’ role is not to form new episodic memories but to transmit semantic memory — accumulated knowledge — to the group. Human elders in indigenous cultures hold the same role. Memory is always doing what it is supposed to do, for the ecology it is embedded in.

AI, memory, and the limits of analogy

Ranganath is cautious about drawing too direct a line between human memory and AI systems. The human solution to the stability–plasticity dilemma (when to update existing knowledge versus preserve it) is the dual system of episodic and semantic memory: episodic memory captures singular events without overwriting general knowledge; sleep then allows the hippocampus to replay recent events and integrate them into the semantic store. This architecture gives humans the flexibility to change on a single trial without catastrophic forgetting of accumulated knowledge.

Building this into AI is not technically trivial. The hard questions — when to record an episode, what information to prioritise, when to use the episode and when to discard it — remain open in humans. Ranganath is also sceptical of the claim that current LLMs replicate human intelligence, pointing out that the test is not whether a system produces correct outputs but whether it makes the same characteristic errors and exhibits the same biases that humans do. Those are absent.

Music, curiosity, and the drive to seek

Throughout the conversation, Ranganath draws on his own experience as a musician and self-described ADHD thinker to illustrate wider claims about memory. The prediction errors that attention-rich brains hunt for in music — the unexpected chord, the odd time signature — are the same signal that drives curiosity more broadly. Dopamine, he argues, is not a pleasure chemical but a drive signal: it fires on the anticipation of a reward, the uncomfortable gap between what you know and what you want to know. That gap is the engine of learning. Close it too quickly — autocomplete the email, stay inside the echo chamber — and the drive atrophies.

His practical advice distils to: diversify your training data. Seek novelty, tolerate dissonance, stress-test your memories by testing yourself rather than re-reading. The brain is not designed to remember everything; it is designed to remember what matters, and what matters is determined by what surprises it.

See also

• Charan Ranganath — speaker
• Lex Fridman — host
• Daniel Kahneman — experiencing self vs remembering self distinction discussed at length
• Elizabeth Loftus — pioneer of false-memory research, cited on misinformation effects
• Why We Remember — Ranganath’s book, discussed throughout
• False Memory — concept page synthesising this episode with Julia Shaw on False Memory, Eyewitness Testimony, and the Psychology of Evil