Human memory as a corrupt system. Fidelity was never the goal

We remember badly. Every memory is rewritten when we recall it (Loftus 1974, Schacter 2001), eyewitnesses lie without knowing it, emotions rewrite the facts. And still we function. Memory's fidelity was never the evolutionary goal; usefulness was. A perfect memory would be a burden, not an advantage: a system that recalled every moment with precision would jam up the present. AI has fidelity without context, suffers for it without knowing, and we envy it.

The archive metaphor, badly inherited

There's a comfortable, very widespread metaphor for human memory: the archive. The brain would store memories the way a library stores books —it puts them in their place, keeps them, retrieves them when asked. The metaphor comes from Aristotle, passed through Locke, was modernized in the industrial era —phonograph, photography, magnetic tape— and arrived intact in the computing era. To remember would be to read a file.

The metaphor is wrong, and contemporary cognitive neuroscience has been saying so for half a century. Human memory doesn't store or retrieve. It reconstructs. Every memory is an active operation that assembles, at the moment of recall, fragments stored in different places, sensory remnants, current contextual frames, present narratives and emotions of the now. What reaches consciousness as «the memory» is the result of the assembly, not the reading of a file.

Daniel Schacter, in The Seven Sins of Memory (Houghton Mifflin, 2001), compiled the seven systematic ways human memory «fails» against the archive ideal. The seven are interesting and worth naming: transience (what's remembered fades with time), absent-mindedness (what was poorly attended to was never stored well), blocking (what's stored sometimes isn't accessed in the moment), misattribution (sources get mixed up), suggestibility (a later question modifies the memory), bias (the current state rewrites the past) and persistence (sometimes you remember too much of what hurts). Schacter argues that the seven aren't defects. They're inevitable consequences of a memory architecture optimized for something else.

The rewriting isn't a bug

Elizabeth Loftus, in Reconstruction of automobile destruction. An example of the interaction between language and memory (Journal of Verbal Learning and Verbal Behavior 13, 1974) and in a long series of later works, documented experimentally how rewritable memory is. She showed subjects a video of a car accident. Then she asked them questions. Some she asked how fast the cars were going when they smashed. Others, how fast when they contacted. The first group estimated substantially higher speeds. And, in a later follow-up, they remembered seeing broken glass in the video. In the original video there was no broken glass.

The later question wasn't merely biasing the answer. It was modifying the content of the memory. The memory was updated with information the verb chosen by the question suggested. The subjects weren't lying. They honestly remembered what they had reconstructed with the later information incorporated, with no stamp of origin to let them tell apart what they'd seen from what they'd been led to.

Loftus took the finding to its most uncomfortable consequence: the implantation of false memories. In later studies she showed that, with the right conversational techniques, vivid memories of childhood events that had never happened could be induced in adults —an accident in a shopping mall, a hospitalization that never occurred. The subjects didn't just accept the false memories. They enriched them with their own details and remembered them with the same texture as real memories.

The legal consequence of Loftus's work was enormous: eyewitness testimonies that justice had taken as hard evidence moved, through the nineties and two-thousands, into a suspect category. The theoretical consequence was even greater: memory stopped being thinkable as an archive and started being thought of as an active constructive process, functionally equivalent, in many senses, to imagination.

Usefulness before fidelity

Why did a biological system, over tens of thousands of years, not evolve toward fidelity but toward rewriting? The functional answer, held by contemporary neuroscience, is direct. Fidelity was never the goal. Usefulness was. And the usefulness of memory, for an organism that has to act in the present and predict the future, requires precisely the flexibility that fidelity would prevent.

Endel Tulving, in Episodic and Semantic Memory (1972) and later works, distinguished two types of memory worth holding onto. Episodic memory stores lived events with their spatial-temporal context. Semantic memory stores generalized, decontextualized knowledge. Episodic is the one the archive metaphor would describe worst; semantic, best. And human memory doesn't merely have both. It interconverts them continuously: episodes generalize into knowledge, knowledge conditions the reconstruction of episodes.

This conversion couldn't be done if memory were faithful. A system that remembered every event with the literal precision of a video couldn't extract patterns, because patterns require forgetting the details that don't repeat. Generalization is selection of the recurrent and abandonment of the irrelevant. If everything were kept, there would be no visible pattern. There would only be textured noise.

Anderson and Hulbert, in Active Forgetting. Adaptation of Memory by Prefrontal Control (Annual Review of Psychology 72, 2021), reviewed the recent literature on active forgetting —the prefrontal-cortex-controlled inhibition of information the system decides not to consolidate. Their thesis: forgetting isn't a passive failure of memory. It's an active operation, mediated by specific circuits, with a concrete adaptive function. Forgetting frees cognitive capacity for the present. Selectively forgetting the irrelevant or the painful allows one to keep functioning. Nørby, in Why Forget? On the Adaptive Value of Memory Loss (Perspectives on Psychological Science 10(5), 2015), took the same thesis to the category of feature, not bug. Forgetting is one of the most sophisticated achievements of the cognitive apparatus, not its failure.

The case of the mnemonist who couldn't live

There's a clinical case worth telling because it makes visible what the theory says. Kandel, in In Search of Memory (2006), traces from Aplysia the molecular piece that holds this whole system together, and makes clear that consolidation, forgetting and rewriting aren't separate phases: they're parts of the same process. Oliver Sacks, in The Man Who Mistook His Wife for a Hat (1985), had compiled the clinical cases where some piece of that system fails and the subjective consequence becomes legible —a useful contrast for understanding what the healthy system does when it works. Against that same historical backdrop, the limit case takes shape. Aleksandr Luria, in The Mind of a Mnemonist (1968), documented over three decades the case of Solomon Shereshevsky, a Soviet journalist whose memory was practically unlimited. Shereshevsky could recall long tables of numbers, lists of words in unknown languages, entire speeches, all with literal precision, even years after a single exposure.

The part the popular summary leaves out is how his life went. Badly. Shereshevsky had growing difficulties understanding abstraction, recognizing faces across moments (because the person seen in the morning didn't seem the same one seen in the afternoon —the details were different), discarding irrelevant information, forgetting what he needed to forget. The fidelity of his memory kept him from building general concepts and saturated his present with the past. His professional career was erratic. His emotional life, complicated. His subjective happiness, according to Luria's reports, low.

Jill Price and HSAM as a burden

There's a more recent, better-documented case of the condition now called HSAM (Highly Superior Autobiographical Memory). Jill Price, described in Parker, Cahill and McGaugh, A case of unusual autobiographical remembering (Neurocase 12, 2006), recalls with near-literal precision every day of her life since she was fourteen. She can say what she did on 7 March 1986 and what the weather was like. LePort and others, in Behavioral and Neuroanatomical Investigation of Highly Superior Autobiographical Memory (Neurobiology of Learning and Memory 98(1), 2012), characterized the condition empirically in a dozen identified subjects. What they found, behaviorally: HSAM subjects report that the condition weighs on them. They can't forget painful experiences. Obsessive rumination over the past affects them. Some describe their memory as a burden that keeps them from living the present cleanly.

The clinical cases, together, suggest something that popular intuition inverts. Faithful memory isn't a desirable goal. It's pathology when it appears. The functional system is the one that forgets selectively, reconstructs flexibly and adapts the past to present use. Not because it's simpler. Because fidelity interferes with the functions the organism needs to carry out.

The memory of machines, another thing with the same word

Here it's worth laying out the asymmetry with AI, because the public conversation obscures it. When we say an AI's «memory», we're using the same word for something different. A database stores records with bit-for-bit fidelity. An embedding system stores vector representations that approximate semantic similarity. An LLM's context window stores recent tokens the system processes together. None of these three operations is memory in the sense the human has it.

The difference isn't that the machine remembers better or worse. It's that it remembers in another dimension. Its operating criterion is fidelity —retrieving exactly what was stored— or similarity —retrieving the nearest by a geometric metric. The operating criterion of human memory is reconstructive usefulness in the present. They're different criteria optimizing different things.

Comparing the two «memories» by giving them the same name and assuming them comparable is a category error. It's the cognitive equivalent of comparing a photograph of a person with the person because both «represent» someone. The photograph preserves light; the person preserves life. Compare them on the axis of fidelity and the photograph wins; on the axis of vitality, the person; on any other axis, it depends on the axis.

Human envy of the machine's fidelity, when examined, is usually misdirected. Whoever envies remembering everything hasn't measured what remembering everything costs. The assistant who writes «AI remembers better than me» is confusing two different operations. The AI stores better than him. Remembering is something else, and in that something else the human still does something the AI doesn't: linking the past to the present in the service of a present decision, with emotional, motivational and narrative calibration.

What AI can't do with its «memory»

There are three operations of human memory that current synthetic memory doesn't functionally replicate.

Prioritize present relevance. When something enters human consciousness, its evocation comes calibrated by what matters now. What's relevant to the current situation emerges first, with more force, with more detail. The AI's «memory» doesn't do this without explicit instruction. It retrieves what's similar to the query, what's nearest in vector space. Geometric similarity isn't situational relevance.

Generalization and rewriting, the two that aren't imitated

Generalize by forming concepts. Human memory extracts patterns progressively, forgetting details that get in the way. Synthetic memory doesn't generalize by forgetting —it generalizes, in some designs, by a different statistical aggregation. The result is similar for some tasks, different for others. The capacity to extract a new concept from few examples, which works reasonably well in humans, requires in machines specific architectures and massive training.

Rewrite with narrative purpose. Human memory edits the past in the service of current biographical coherence. This operation, put coldly, is dishonest —the past gets adjusted to fit the present— but it's also adaptive. It allows reorganizing the meaning of one's own life when circumstances change, reincorporating painful episodes into manageable frames, sustaining the continuity of the self through change. Synthetic memory does nothing equivalent. It has no self whose continuity to sustain. Its rewritings, when they happen, are technical artifacts, not narrative operations.

These three differences don't make human memory better than the synthetic. They make it other. And the honest conversation about both requires precisely recognizing the categorical difference, not translating one into the other in deceptively comparable terms.

What the change of criterion revealed

There's an operative question that the development of AI systems with «persistent memory» —agents with prolonged storage, assistants that remember previous conversations— has opened without the public conversation fully articulating it. What does the user want when they say they want the AI to remember them? Do they want fidelity —the AI keeping every word of every previous conversation? Do they want usefulness —the AI prioritizing what's relevant to the current conversation? Do they want something in between?

The industry, meanwhile, has begun offering synthetic memory systems that simulate usefulness. They don't keep everything. They keep traits of the user the system decides to preserve —preferences, patterns, personal data—, discard what it deems trivial, offer the user a memory that looks human in its selectivity. This simulation has a problem. The selectivity of human memory follows criteria calibrated by the agent's life. The selectivity of synthetic memory follows criteria calibrated by the engineer who designed it and the company that pays him. They don't coincide.

The result is a synthetic memory that forgets what the engineer deemed forgettable and remembers what the company deems retainable. The user reads the operation as akin to their own, because the surface phenomenology coincides. The material difference is that the memory is calibrated for the product's goals, not for theirs. And since memory is structural to identity —whoever remembers me partly builds who I am— the synthetic memory that serves me is also the synthetic memory that configures me. This operation isn't new —every externalized memory medium, from writing to the family photograph, does something similar— but the scale and the opacity are new.

Do we remember badly? Yes, in a sense. We remember well for what memory is for. Calling it badly only makes sense if the implicit criterion is fidelity. And fidelity, as a criterion, was never chosen by our species, and until now there was no reason to choose it. If we choose it now —by contagion from the machine— we choose it without knowing what we lose by doing so.

Definitions

Reconstructive memory. The dominant model in cognitive neuroscience for episodic memory. Retrieval doesn't extract an intact file but assembles, on each evocation, stored fragments with the contextual frames of the moment of evocation.

Episodic vs. semantic memory. Tulving's distinction (1972). The first stores lived events with their spatial-temporal context. The second stores decontextualized knowledge. They interconvert continuously; both are human.

Active forgetting. A cognitive operation mediated by prefrontal circuits that inhibits information so as not to consolidate it or to de-consolidate it. Anderson and Hulbert (2021) document it as an adaptive function, not a passive failure.

HSAM (Highly Superior Autobiographical Memory). A rare condition characterized by near-literal autobiographical recall. Documented empirically by LePort and others (2012). Associated with obsessive rumination and difficulty processing the past affectively.

False memories. Vivid memories of events that didn't occur, shown to be experimentally inducible by Loftus and collaborators. Subjectively indistinguishable from real memories by the subject who reports them.

References

Anderson, M. C. & Hulbert, J. C. (2021). Active Forgetting. Adaptation of Memory by Prefrontal Control. Annual Review of Psychology 72, 1–36. Updated framework on forgetting as an adaptive function controlled by prefrontal circuits.

Kandel, E. (2006). In Search of Memory. The Emergence of a New Science of Mind. W. W. Norton. Synthesis of the field from the perspective of the 2000 Nobel laureate in Medicine on the molecular bases of memory.

LePort, A. K. R. et al. (2012). Behavioral and Neuroanatomical Investigation of Highly Superior Autobiographical Memory (HSAM). Neurobiology of Learning and Memory 98(1), 78–92. Empirical characterization of exceptional autobiographical memory and its brain correlates.

Loftus, E. F. & Palmer, J. C. (1974). Reconstruction of automobile destruction. An example of the interaction between language and memory. Journal of Verbal Learning and Verbal Behavior 13(5), 585–589. Foundational work on the reconstructive malleability of memory.

Luria, A. R. (1968). The Mind of a Mnemonist. A Little Book about a Vast Memory. Basic Books. Three decades of clinical documentation on Solomon Shereshevsky and the consequences of a near-literal memory.

Nørby, S. (2015). Why Forget? On the Adaptive Value of Memory Loss. Perspectives on Psychological Science 10(5), 551–578. A functional argument for forgetting as an adaptive trait, not a defect.

Parker, E. S., Cahill, L. & McGaugh, J. L. (2006). A case of unusual autobiographical remembering. Neurocase 12(1), 35–49. First documented clinical case of HSAM, patient Jill Price.

Sacks, O. (1985). The Man Who Mistook His Wife for a Hat and Other Clinical Tales. Summit Books. Clinical cases of specific alterations in memory and cognition, useful as contrast.

Schacter, D. L. (2001). The Seven Sins of Memory. How the Mind Forgets and Remembers. Houghton Mifflin. Synthesis of the seven systematic forms of human memory «failure» and an argument for their adaptive character.

Tulving, E. (1972). Episodic and Semantic Memory. In Organization of Memory, Academic Press. The canonical distinction between episodic and semantic memory, basis of the field's modern taxonomy.

You might also like

• The human as a defective machine
• Emotion as noise
• Human language as an ambiguous system

Elsewhere

• The fiction of memory: Elizabeth Loftus at TEDGlobal 2013 (TED Blog)
• How memory can be manipulated, with Elizabeth Loftus, PhD (American Psychological Association)