Episodic vs Semantic Memory: The Difference Between Knowing Paris and Remembering Paris

Introduction

A man in Toronto could tell you what a motorcycle is. He could explain how the engine works, name the parts, describe the difference between a road bike and a dirt bike. He knew that he had owned one. He could not remember a single ride.

His name was Kent Cochrane, and in the memory literature he is known as K.C. In October 1981, at the age of thirty, he came off his motorcycle and sustained extensive damage to both medial temporal lobes. What the accident left behind was one of the most precisely shaped deficits ever documented. His intelligence was intact. His vocabulary was intact. His knowledge of the world, including facts about his own life, was largely intact. What had gone was the ability to re-experience anything that had ever happened to him. He was studied for roughly three decades across more than twenty papers, and when he died in 2014 his brain was examined directly [1].

K.C. is the cleanest illustration of an idea that has organised memory research since 1972. Some of what you carry is knowledge. Some of it is experience. The first kind has no timestamp and no viewpoint. The second is anchored to a moment and comes with the peculiar sensation of being there again.

Endel Tulving called the first semantic memory and the second episodic memory [2]. The distinction became one of the most useful ideas in cognitive psychology. It also became one of the most argued-over, and that second half of the story is almost entirely missing from the places people go to look it up.

This article covers both halves. What the distinction is, what the evidence for it looks like, how researchers actually measure it, and why fifty years of work has left the field genuinely divided about whether these are two systems or one thing seen from two angles.

One thing to hold onto before starting. Several of the studies below involve very small numbers of people. Three patients here, eleven there, thirteen infants in one case. Small samples are how this field advances, because the conditions being studied are rare. They are also a reason to hold conclusions loosely. Every sample size in this article is stated for exactly that reason.

Knowing Paris and Remembering Paris

Start with the example in the title, because it does the work faster than a definition.

You probably know that Paris is the capital of France. You know it sits on the Seine, that the Eiffel Tower is there, that people speak French. You did not learn any of that at an identifiable moment. There is no scene attached. If someone asked when you first learned Paris was the capital of France, you would have nothing to offer. The fact arrived from somewhere and stayed, and the delivery vehicle dissolved.

That is semantic memory. General knowledge, stripped of the circumstances in which you acquired it.

Now suppose you have been to Paris. Something else is available. A particular afternoon, a particular street, whatever the weather was doing, the specific and slightly disappointing sandwich. You are not retrieving a fact about that afternoon. You are, in some limited and imperfect way, back in it.

That is episodic memory. A specific event, located in time and place, retrieved from your own point of view.

Tulving argued that the difference between these is not just a difference in content. It is a difference in the kind of consciousness involved. Retrieving a fact feels like knowing. Retrieving an episode feels like remembering, and remembering has a first-person quality that knowing does not. He gave these two states names in 1985: noetic consciousness for knowing, autonoetic consciousness for the self-aware re-experiencing that accompanies episodic retrieval [3].

Autonoetic is the load-bearing word in the whole framework. It is also, as the later sections will show, the one that causes most of the trouble, because you cannot observe it from outside. You can only ask.

That table is a summary, not a verdict. Almost every row in it has been argued about, and the sections that follow explain which arguments are still open.

A Century Before Anyone Had a Name for It

The idea that memory comes in more than one variety is older than the terminology.

Hermann Ebbinghaus began the experimental study of memory in 1885, memorising lists of nonsense syllables on himself and plotting how fast they disappeared. His concern was the shape of forgetting rather than the kinds of memory, but he established that memory could be measured at all. Eleven years later, in 1896, Henri Bergson drew a philosophical distinction in Matter and Memory between habit-memory, the kind built by repetition, and what he called pure or representational memory, the recollection of singular events. Frederic Bartlett showed in 1932 that remembering is reconstructive rather than reproductive, that people rebuild a story each time from schemas rather than replaying a recording. In the 1950s Wilder Penfield reported that electrical stimulation of the temporal lobe during surgery could evoke what patients described as vivid experiential recall.

Then came the case that reorganised the field.

In 1957 William Scoville and Brenda Milner published nine cases of bilateral medial temporal-lobe resection [4]. The most severely affected, known for decades only as H.M. and identified after his death in 2008 as Henry Molaison, had undergone a resection extending roughly eight centimetres back from the temporal tips, taking with it approximately the anterior two-thirds of the hippocampus and hippocampal gyrus. He could no longer form new memories of his own experience.

Two details from that paper matter more than they are usually given credit for. Removing only the uncus and amygdala bilaterally did not produce memory impairment. And the severity of the amnesia tracked how much hippocampus had been removed. The hippocampus was not incidental. It was the variable.

Fifteen years later, in a chapter in the edited volume Organization of Memory, Tulving proposed the split that the rest of this article is about. He suggested that episodic and semantic memory were not two labels for one underlying store but two functionally distinct systems.

Notice the shape of that list. Seventy-two years separate the first measurement of memory from the naming of the distinction. Fifty more years separate the naming from a journal issue devoted to questioning it. This is not a field that settles things quickly.

The Evidence That Made the Distinction Stick

The reason the split survived is that brain damage does not respect tidy categories unless the categories are real. Three lines of evidence carried most of the weight.

The first is K.C. himself. Profound episodic amnesia, covering both new events and personal events from before the accident, alongside largely preserved semantic memory and preserved implicit memory. He knew facts about his own history without being able to revisit any of it.

The second is developmental amnesia. In 1997 Faraneh Vargha-Khadem and colleagues reported three patients whose bilateral hippocampal pathology arose at birth, at age four, and at age nine [5]. Three patients is a small sample, and the finding is striking enough that the number should stay visible. All three had normal IQ. All three progressed through mainstream schooling. All three acquired semantic knowledge, the ordinary factual content of an education, despite severe episodic impairment. On delayed recall they performed in the range of twenty to twenty-five percent of controls, whether the task was retelling a story or redrawing a geometric design.

Children who cannot remember their lessons, learning the content of those lessons anyway. If episodic and semantic memory were the same system, that should not be possible.

The third is the mirror image. Semantic dementia, described in detail by John Hodges and colleagues in 1992, involves progressive atrophy of the anterior temporal lobe [6]. Patients lose the meanings of words and the identities of objects while their memory for recent personal events remains comparatively intact. Alzheimer's disease characteristically runs the other way, with episodic impairment appearing first. Two conditions, opposite profiles. In neuropsychology that pattern is called a double dissociation, and it is the strongest argument the two-systems view has.

Anyone wanting the full picture of the knowledge side of this pairing will find it in the dedicated treatment of how semantic memory is organised, and the clinical mirror is covered in the account of what semantic dementia takes and what it leaves.

The Brain Basis, and Where It Gets Complicated

The anatomy has two centres of gravity.

The hippocampus binds the elements of an event together into a trace. Sight, sound, place, sequence and the sense of a viewpoint are held apart in different cortical regions, and something has to tie them into one retrievable object. Detailed recollection depends on that binding operation, which is why hippocampal damage takes episodic memory first. The mechanism is covered more fully in the account of how the hippocampus decides what to keep.

Two microcircuit properties matter here. The dentate gyrus supports pattern separation, keeping similar events from collapsing into each other, so that this Tuesday's commute does not overwrite last Tuesday's. Area CA3 supports pattern completion, reconstructing a whole memory from a partial cue. Separation is what makes episodes distinct; completion is what makes a fragment enough to bring one back.

Semantic memory has a different architecture. Karalyn Patterson, Peter Nestor and Timothy Rogers proposed the hub-and-spoke model in 2007 [7]. Modality-specific regions, the spokes, handle what things look like, sound like and are used for. These converge on a transmodal hub in the anterior temporal lobe that represents meaning independently of any single sense. Matthew Lambon Ralph and colleagues developed the computational account of this in 2017 [8]. On this model, semantic dementia is what happens when the hub degrades, which explains why the deficit crosses every modality at once rather than affecting only words or only pictures.

Charan Ranganath and Maureen Ritchey added a systems-level framing in 2012, distinguishing a posterior-medial network concerned with context and situations from an anterior-temporal network concerned with entities and items [9]. That framing is useful precisely because it does not map cleanly onto episodic versus semantic. It cuts the space differently.

Here is where the tidy story starts to come apart. The networks overlap substantially. Louis Renoult, Muireann Irish, Morris Moscovitch and Michael Rugg, reviewing the neuroimaging evidence in 2019, concluded that episodic and semantic memory are "inextricably intertwined, yet retain a measure of distinctiveness" [10]. Two systems that share most of their machinery are harder to call two systems.

How Researchers Actually Tell Them Apart

This is the section that almost nobody writes, and it is the one that determines how much weight the rest deserves. A distinction is only as good as the instruments that measure it.

The Autobiographical Interview, introduced by Brian Levine and colleagues in 2002, is the workhorse [11]. A participant recalls a personal event freely, and trained scorers divide the resulting narrative into internal details, which are episodic and specific to that event, and external details, which are semantic, repeated, or drawn from elsewhere. It is widely used, and it produced one of the most replicable findings in cognitive ageing: older adults generate fewer internal details and more external ones.

It also has limits worth stating plainly. The internal and external categories each bundle together several different kinds of content, so the split is coarser than it looks. And because the measure operates on free narrative, a person who simply talks more, or talks in a more discursive style, can shift their own scores without any underlying difference in memory.

The Remember/Know procedure came out of Tulving's 1985 paper. A participant who recognises an item says "Remember" if they can bring back something about the encounter with it, and "Know" if it merely feels familiar. It is elegant and it has a serious complication. Andrew Yonelinas, reviewing thirty years of work in 2002, argued that recollection behaves like a threshold process that either delivers qualitative detail or does not, while familiarity behaves like a continuous strength signal [12]. On that reading, Remember/Know measures two retrieval processes rather than two memory systems. The related asymmetry between recognising something and producing it from nothing is explored in the piece on recognition and recall.

The Autobiographical Memory Interview, developed by Michael Kopelman, Barbara Wilson and Alan Baddeley in 1989, takes a different approach [13]. It scores autobiographical incidents and personal semantic facts separately across defined life periods, which is what makes it able to document temporally graded retrograde amnesia.

On the semantic side, the Pyramids and Palm Trees Test, published by David Howard and Karalyn Patterson in 1992, asks a participant to match a pyramid with a palm tree rather than a pine tree. It is a test kit rather than a journal article, and it works nonverbally, which is what makes it useful when language itself is impaired. The Camel and Cactus Test is a related semantic-association measure used with semantic dementia patients.

Verbal fluency splits usefully too. Category fluency, naming as many animals as possible in a minute, is more sensitive to anterior temporal and semantic degradation. Letter fluency, naming words beginning with F, loads more heavily on frontal and executive processes. The gap between a person's two scores carries diagnostic information.

Finally, the Survey of Autobiographical Memory, a trait self-report developed by Daniela Palombo and colleagues in 2013, asks people to rate their own episodic, semantic, spatial and future-thinking memory. It is how one of the most interesting conditions in this article was identified.

Notice what has happened across this section. Every instrument measures something slightly different, and several of them measure processes rather than systems. That is not a flaw in the science. It is the reason the underlying question stayed open.

From Episode to Knowledge

If the two are distinct, something has to connect them, because knowledge has to come from somewhere. Nobody is born knowing that Paris is the capital of France.

The dominant account is complementary learning systems, published by James McClelland, Bruce McNaughton and Randall O'Reilly in 1995 [14]. The argument starts from a problem in artificial neural networks that Michael McCloskey and Neal Cohen had identified in 1989: train a single network on new material and it overwrites the old, a failure known as catastrophic interference.

Brains do not do this. The proposed reason is that they use two learning systems with opposite settings. The hippocampus learns fast, using sparse and well-separated codes, which lets it capture a single event in one exposure without disturbing its neighbours. The neocortex learns slowly, using distributed and overlapping codes, gradually extracting what is common across many episodes. Slow learning is not a limitation here. It is what allows structure to emerge from repetition instead of the last example simply replacing the previous one. Dharshan Kumaran, Demis Hassabis and McClelland updated the framework in 2016 [15].

The transfer happens substantially during sleep. In 1994 Matthew Wilson and Bruce McNaughton recorded hippocampal place cells in rats and found that ensembles which fired together during waking exploration fired together again during slow-wave sleep [16]. The brain was replaying the day. Susanne Diekelmann and Jan Born set out the active systems consolidation account in 2010, in which slow oscillations, sleep spindles and sharp-wave ripples coordinate the redistribution of hippocampal memories to neocortex [17]. The overnight side of this is covered further in the piece on what the sleeping brain does with the day.

Prior knowledge speeds the whole process dramatically. Dorothy Tse and colleagues showed in 2007 that rats with an established flavour-place schema could make new paired associates hippocampus-independent within roughly forty-eight hours, rather than the weeks that consolidation normally takes [18]. That is a rat study, and the mechanism should not be assumed to transfer to humans unchanged. But it points at something familiar: learning is faster in a subject you already know, and it is faster because the framework is already there for the new material to attach to.

There is a serious disagreement buried in all of this, and it is covered in the next section but one.

That chart is a schematic drawn from the Betula longitudinal study, and it should be read as a shape rather than as a set of published data points. Michael Rönnlund, Lars Nyberg, Lars Bäckman and Lars-Göran Nilsson followed 829 adults and reported two things that had been obscured in earlier cross-sectional work [19]. Episodic memory showed no longitudinal decrement before age sixty, even after adjusting for practice effects. Semantic memory showed small increases until around age fifty-five, and then declined more gently in old age than episodic memory did.

The line is episodic, the bars are semantic, and the point is the divergence rather than the exact values. The popular belief that episodic memory starts sliding in your twenties does not survive contact with longitudinal data. Lars Nyberg and colleagues set out the broader picture of brain maintenance in ageing in 2012 [20], and the fuller treatment is in the article on how memory actually changes with age.

The Grey Zone Nobody Talks About

Ask yourself where you went to primary school.

You know the answer. It is a fact about your own life. But it is not a fact about the world in the way that the capital of France is, because it is true only for you and nobody else shares it. And you almost certainly are not retrieving a scene to answer it. You just know.

Where does that go?

Louis Renoult, Patrick Davidson, Daniela Palombo, Morris Moscovitch and Brian Levine gave this its own name in 2012: personal semantics [21]. It covers autobiographical facts, repeated or generic personal events, and self-knowledge. It is idiosyncratically personal, like episodic memory, and simultaneously decontextualised and fact-like, like semantic memory. Depending on how researchers operationalise it, its neural correlates can resemble the semantic system, the episodic system, or neither cleanly.

Annick Tanguay and colleagues compared personal-semantic, general-semantic and episodic memory directly with neuroimaging in 2023 and found both shared and distinctive correlates [22].

Personal semantics matters out of proportion to its coverage, for two reasons. It makes up an enormous share of what people actually retrieve about their own lives in ordinary conversation. And it sits exactly where a two-category system should have nothing. A clean dichotomy with a well-populated middle is not a clean dichotomy.

The Argument That Has Run for Fifty Years

Here is what almost every explainer on this subject leaves out.

The episodic and semantic distinction is not settled science. It is a live and unresolved argument, and it has been running since shortly after Tulving proposed it.

The case for two systems is Tulving's, developed across 1972, 1983, 1985 and his 2002 review. Episodic memory is defined by more than its content. It involves the self, subjective time, and autonoetic consciousness. On this view the double dissociations are real and the systems are genuinely separate.

The case against is Larry Squire's. Squire and Stuart Zola argued in 1998 that both are components of a single declarative memory system dependent on the medial temporal lobe, and that the evidence for clean differential impairment in amnesia does not hold up under scrutiny [23]. Their strongest point concerns the developmental amnesia cases. Those patients do acquire semantic knowledge, but not normally. Acquisition is slower, and source errors are common. If the semantic system were genuinely untouched, it should not be impaired at all.

That objection has never been fully answered.

A third position concerns consolidation specifically. Standard systems consolidation holds that memories gradually become independent of the hippocampus. Multiple trace theory, proposed by Lynn Nadel and Morris Moscovitch in 1997 and developed into trace transformation theory since, holds that detailed episodic memories always require hippocampal reactivation no matter how old they are [24]. On this account consolidation does not hand a memory over. It produces a semanticised version in neocortex while the episodic trace stays hippocampus-dependent, so that an old memory can be retrieved either as gist or as an episode depending on which representation is accessed. Moscovitch, Roberto Cabeza, Gordon Winocur and Nadel set out the mature version in 2016 [25], and Melanie Sekeres, Winocur and Moscovitch extended it in 2018 [26].

Then there are the positions that reject the binary framing itself. Muireann Irish and Deniz Vatansever proposed in 2020 that episodic and semantic memory sit on a gradient rather than in separate boxes [27]. David Rubin went further in 2022, proposing a three-dimensional conceptual space in which the traditional hierarchy is abandoned altogether and memories are located by their properties rather than sorted into types [28].

In 2022 the journal Memory and Cognition marked the fiftieth anniversary of Tulving's chapter with a special issue whose organising question, in the editorial by Felipe De Brigard, Sharda Umanath and Irish, was whether the distinction should be rethought [29]. A field does not run an anniversary issue asking whether its foundational distinction still holds unless the answer is genuinely in doubt.

Endel Tulving died on 11 September 2023, at the age of ninety-six. The argument he started outlived him and shows no sign of closing.

The Cases at the Edges

The boundaries are where this gets genuinely strange, and where the abstractions become people.

Severely deficient autobiographical memory, or SDAM, was described by Daniela Palombo, Claude Alain, Hedvig Söderlund, Wayne Khuu and Brian Levine in 2015 [30]. Three healthy, high-functioning adults, and three is the entire sample, reported a lifelong inability to relive their own past. Not amnesia. They know what happened to them. They simply cannot re-experience any of it. The investigators found reduced visual memory, reduced right hippocampal volume, reduced activation in the autobiographical network, and a markedly reduced late-positive event-related potential associated with recollection. Recognition memory was intact and daily life was unaffected, apparently supported by semantic knowledge of their own history.

If autonoetic re-experiencing can be absent in someone whose life runs normally, then it is dissociable from the knowledge it usually accompanies.

Aphantasia is the absence of voluntary visual imagery. Adam Zeman and colleagues surveyed 1,288 UK participants in 2020 and reported prevalence figures across the imagery spectrum: 0.7 percent for extreme aphantasia, 2.6 percent for aphantasia, 11.2 percent for hyperphantasia, and 2.6 percent for extreme hyperphantasia [31]. Alexei Dawes and colleagues found that aphantasic participants report reduced episodic and autobiographical detail while semantic memory is relatively spared, with the smallest group difference of all appearing in semantic memory [32]. Zeman reviewed the field in 2024 [33]. The caveat that belongs alongside these numbers is that aphantasia is defined largely by self-report on a questionnaire. Objective findings in priming and pupillometry support that it is a real phenomenon, but the foundation is subjective. The fuller account is in the piece on memory without mental pictures.

Highly superior autobiographical memory runs the other way. Aurora LePort and colleagues studied eleven confirmed HSAM individuals in 2012 and identified nine brain structures that differed morphologically from those of controls [34]. The striking result is what these people were not better at. Their performance on most standard laboratory memory tests was comparable to matched controls. The advantage was specific to autobiographical recall and to dating events. This is not superhuman memory. It is one system, selectively enhanced.

Animals are where the definitional problem bites hardest. In 1998 Nicola Clayton and Anthony Dickinson showed that western scrub jays cached perishable wax worms and non-perishable peanuts and later recovered them in a pattern that tracked what they had hidden, where, and how long ago [35]. That is the operational definition of episodic memory in every respect except one. Nobody can ask a jay whether it re-experienced the caching. Thomas Suddendorf and Michael Corballis argued in 2007 that behaviour alone cannot establish autonoetic consciousness [36]. This is why the literature says episodic-like memory rather than episodic memory, and the hyphenated qualifier is doing real work rather than hedging.

Infancy delivered the most recent revision. Katherine Akers and colleagues proposed in 2014 that high rates of hippocampal neurogenesis in infancy destabilise existing memories, showing in mice that increasing neurogenesis after learning induced forgetting while decreasing it during infancy reduced it [37]. Mice, not humans, and the distinction should be kept firmly in view. Then in 2025 Tristan Yates and colleagues managed awake-infant fMRI and found that greater hippocampal activity while viewing novel photographs predicted later memory-based looking [38]. The effect emerged around age one, holding for 11 of 13 infants older than a year and not for the younger ones. Thirteen infants is a very small study and the finding needs replication. But it points somewhere specific: infant brains do encode. Whatever produces childhood amnesia looks more like a later failure of access than a failure to record. The full account is in the article on why the first years go missing.

Why Episodic Memory Is the Unreliable One

There is a reason episodic memory drifts and semantic memory mostly does not, and it turns out to be a feature rather than a defect.

Daniel Schacter and Donna Rose Addis proposed constructive episodic simulation in 2007 [39]. Remembering the past and imagining the future recruit a substantially overlapping core network, which Donna Rose Addis, Alana Wong and Schacter demonstrated by separating event construction from event elaboration [40]. Their argument is that episodic memory is not a recording system at all. It is a system for flexibly recombining stored elements, and the reason it can be recombined into a future scenario you have never experienced is precisely the reason it can be recombined into a past that did not happen quite that way.

The flexibility and the unreliability are the same property. You cannot have one without the other.

This is also why retrieval changes the memory. Bringing an episode back puts it into a labile state before it is stored again, which means every act of remembering is an opportunity for the trace to shift. The mechanism is set out in the piece on how memories are rewritten when you recall them, and the broader reconstructive account is covered in why memory builds rather than replays.

Semantic memory is comparatively stable because it has been abstracted across many episodes. The capital of France has been confirmed hundreds of times from independent sources. A single Tuesday afternoon has been confirmed once, by one witness, who was not taking notes.

What This Means If You Are Trying to Learn Something

The episodic-to-semantic pipeline is not an abstraction. It is the thing that happens when anyone studies anything, and understanding it explains why some study methods work better than others.

Consider what happens when you learn a new term. The first encounter is episodic. It has a setting, a page, a time of day, possibly a mood. What you want, eventually, is for it to become semantic: available whenever you need it, with no dependence on where you were sitting. The gap between those two states is the whole problem of studying.

Two well-established effects operate directly on that gap.

The testing effect is that retrieving information strengthens it more than re-reading does. In the framework of this article, retrieval practice repeatedly reactivates a trace and drives the abstraction that turns a specific encounter into general knowledge. Re-reading gives you fluency with the page, which is a property of the episode, and fluency with the page is easy to mistake for knowing the material. The mechanism is covered in the article on why retrieval beats review.

The spacing effect is that the same total study time produces more durable knowledge when it is distributed rather than massed. Spacing gives consolidation, including sleep-dependent replay, the opportunity to run between sessions. The evidence is set out in the treatment of why spacing outperforms cramming.

There is a third point that follows from Tse and colleagues. Because prior knowledge accelerates consolidation, learning is genuinely easier in a domain where you already have a framework. The practical implication is that time spent building a rough structure first is not time lost before the real learning starts. It is what makes the later material stick faster.

One more thing follows from the encoding side. Episodic memory is bound to context, which is why the room, the time of day and even your mood at encoding can become retrieval cues, an effect explored in the piece on why place affects recall. Knowledge that has become properly semantic has shed those dependencies. That shedding is the goal.

What Is Settled and What Is Not

It is worth separating these explicitly, because the confident tone of most summaries obscures the line.

Reasonably settled. Tulving's 1972 proposal reorganised memory research. Hippocampal and medial temporal damage impairs memory for personal events far more than it impairs general knowledge. Semantic dementia and Alzheimer's disease produce opposite clinical profiles. Episodic and semantic memory follow measurably different trajectories across adult life. Sleep-dependent replay contributes to consolidation. Prior knowledge accelerates consolidation, at least in rats. Remembering the past and imagining the future share a core network. Episodic memory is more distortion-prone than semantic memory.

Genuinely unresolved. Whether these are two systems or two regions of one continuum. Whether the double dissociations are as clean as they are usually presented, given that developmental amnesics do show semantic learning deficits. Whether detailed episodic memories ever become independent of the hippocampus. Whether any non-human animal has autonoetic consciousness, or whether the question is answerable at all. How much weight self-report can carry in aphantasia research. Whether the Remember/Know procedure measures systems or processes.

That second list is not a weakness in the field. It is the field. The distinction has been productive for fifty years precisely because it generated questions sharp enough to argue about.

Conclusion

Return to K.C. one last time.

He knew he had owned a motorcycle. He could not remember riding one. Both statements were about the same machine, the same person, the same life. Only one of them was available to him.

That gap is what Tulving named in 1972, and it is why the distinction has lasted despite everything that has been thrown at it. Something about knowing that Paris is the capital of France is different in kind from remembering an afternoon there, and any account of memory that cannot express the difference is missing something a brain injury can reveal in an afternoon.

What fifty years have added is caution about how sharp the line is. The networks overlap. Personal semantics occupies the middle. The measurement tools capture processes as often as systems. The strongest counterargument, that these are components of one declarative system, has never been decisively refuted.

The honest position is that the distinction is a good description that has not resolved into a clean mechanism. It survives because it keeps predicting things: which patients will show which deficits, which abilities will decline together in ageing, why a fact and an event feel so different from the inside. It stays contested because prediction is not the same as explanation.

Tulving spent five decades on the difference between knowing and remembering. The most useful thing he left behind may be the argument itself.