How memory works
Episodic and semantic memory, explained
Episodic memory is your memory for things that happened to you, with a when and a where. Semantic memory is what you know about the world, without needing to recall how you learned it. They’re two kinds of long-term memory that work together nearly all the time.
You know that Paris is the capital of France, and you can’t say when you learned it. You remember the evening your train left without you there, and you’re back on the platform as you think of it.
- Named by
- Endel Tulving, 1972; “semantic memory” was already in use
- Episodic holds
- events from your own past, and when and where they happened
- Semantic holds
- facts, meanings and concepts, with no record of how you got them
- Both are
- long-term and declarative: things you can state
- Classic test
- remember or know, from Tulving, 1985
An AI drafted this page from the sources below and another AI checked it. No person has tried the demos yet.
Try it
Do you remember it, or do you just know?
When you recognise something, there can be two different feelings. This is a small version of the task Tulving introduced in 1985 to get at that.
About three minutes. First you see 16 words, each with a quick question: some about the thing’s size, some about the word’s spelling. Just answer the question. After a short pause you see 32 words, and half are new.
For each one, you say which of these is true.
- Remember: something comes back from when you saw it: what you pictured, what came just before, what you thought.
- Know: you’re fairly sure it was there, and nothing about seeing it comes back.
- New: it wasn’t on the list.
It’s a demonstration. Kihlstrom’s version below tested 160 words, half of them new, a day later, and asking people to judge their own experience has limits. Your answers stay on this page, and we only count that a demo was started.
If you can’t see it: words appear one at a time with buttons, and a screen reader should announce each word. The task can also be done by a helper reading out a list: ask a quick question about each word, wait, then read a mix of old and new words, and say for each one whether you remember it, just know it, or think it’s new.
What’s the difference between episodic and semantic memory?
This is Tulving’s 1972 contrast, as Renoult and Rugg report it, with what’s happened to each point since.
| Episodic | Semantic | Since then | |
|---|---|---|---|
| Holds | “personally experienced unique events” | a “mental thesaurus”: words, meanings, rules | Semantic memory widened to “knowledge of the world”, not all of it verbal |
| Refers to | your own past | the world, “detached from autobiographical reference” | Still accepted, and Renoult and Rugg call it uncontroversial for episodic memory |
| Record of how you got it | yes | no | Researchers still have “little experimental control” over how semantic knowledge is acquired |
| Stores | how things looked and sounded | “cognitive referents” | Current views root semantic memory in perception and action too |
| Forgetting | more prone to it | less prone to it | Tulving later called single events more fragile. Renoult and Rugg say a fair comparison is hard to run |
| Changed by recalling | yes | no | The idea that recall changes episodic memories now has much research behind it |
Try sorting some
Most examples sort easily. The last two are in here on purpose.
Nine memories to sort, a minute or two. The explanations matter more than the score.
Where did the distinction come from?
Tulving helped run a conference on memory in Pittsburgh in 1971. Three sets of speakers talked about a newly named thing, semantic memory, first described by Quillian in 1966. Tulving later wrote that they “were not really talking about the kind of memory” psychologists had studied since Ebbinghaus.
He went home and wrote the chapter in reaction. He needed a name for the other kind, and had misgivings about “event memory”. He then came across “nonepisodic memory” in a small book by S. Munsat, and took the positive version.
In 1972 he presented it as a “pre-theoretical position”, and Renoult and Rugg say this first version made no strong claim for two separate systems. The claim that one system “can operate independently of the other” came in his 1983 book.
The feeling came later
The 1972 version was about what each kind stores and refers to, and how retrieval and forgetting affect it. By 1983 Tulving was calling the “subjective awareness of remembering” the thing that had to be explained. In 1985 he named it autonoetic awareness, meaning self-knowing: the sense of being back in your own past, which he also called mental time travel.
The test
A feeling is hard to study, so Tulving asked people directly: when you recognise a word, do you remember seeing it, or do you just know it was there? Gardiner’s 1988 version added a question at study, about sound or about meaning. As Kihlstrom reports it, thinking about meaning raised “remember” answers and left “know” answers about where they were.
The same numbers as a table
| Word | Remember | Know | Feel |
|---|---|---|---|
| Generated an associate at study (meaning) | 0.37 | 0.25 | 0.19 |
| Generated a rhyme at study (sound) | 0.27 | 0.22 | 0.26 |
| New words, wrongly called old | 0.05 | 0.07 | 0.23 |
The chart comes from Kihlstrom’s 2019 replication, since Gardiner’s own paper wasn’t available to us, and it complicates the story. Pool know and feel into Gardiner’s single “know” and it stays about flat, 0.48 to 0.44, which repeats his result. Split them, and “remember” rose with meaning, “feel” fell, and “know” rose slightly.
Taking remember and know alone, they responded to the study task in much the same way: the difference between them wasn’t statistically significant. Across his three experiments they also matched on confidence, accuracy and speed, and Kihlstrom says this leaves their distinction “somewhat in doubt”.
How do they work together?
From the first, Tulving called the two “partially overlapping”, and by 1983 “closely interdependent”, interacting “virtually all the time”. Here are two kinds of traffic between them.
- Knowledge shapes events. Episodes almost always have meaning in them, so laying one down usually draws on what you already know. Tulving wrote in 1985 that recollection is “determined jointly by episodic and semantic information”.
- Knowledge can start out tied to an event. In Conway and colleagues’ studies of learning new facts, the most common answer on a first test was remembering the circumstances, and on a later test it was “just knowing”. They call this the R-to-K (remember-to-know) shift.
You may have noticed something like the second one yourself: a fact from a lecture is tied to the lecture for a while, and then it’s just something you know. How that happens isn’t settled, and the Field Guide’s page sets out the argument over whether gist comes from compressed episodes or is laid down alongside them.
Is episodic memory a recording?
Tulving’s 1972 chapter called it “a more or less faithful record”. By 1983 he had added that “a good deal” of remembering is not veridical. The field has moved further since: Renoult and Rugg report some authors’ view that episodic memories are “summarized and generic”, and often hold “only the gist of an episode”.
Schacter and Addis ask why it would work this way, in contrast to a video recorder. Their hypothesis is that piecing memories together from fragments lets you do something else with the same machinery: imagine what hasn’t happened yet. Since the future is never an exact repeat, it helps to have a system that can recombine the past.
Where are they in the brain?
Tulving argued in 1983 that the two could be told apart anatomically, citing studies of brain damage that affects one more than the other. Those cases are much of the classic evidence for the split. One well-studied example is K.C., the initials the papers use for a man left with widespread brain damage, including large lesions of the hippocampus on both sides, by a motorcycle accident. Rosenbaum and colleagues report that he kept a good deal of his general knowledge of the world, including knowledge about himself, and couldn’t recollect any event from his own life. Renoult and Rugg now write of “extensive overlap in their neural substrates”, and say the two keep “a degree of distinctiveness”.
- Episodic: Simply Psychology’s summary puts the medial temporal lobe, including the hippocampus, at the centre. Wheeler, Stuss and Tulving argued in 1997 that the prefrontal cortex plays a key part too, especially in the felt sense of remembering.
- Semantic: Binder and Desai’s review finds both sensory and motor systems at work and large convergence regions, including the inferior parietal lobe and much of the temporal lobe. It reports a meta-analysis of 120 imaging studies.
How do they fit with other kinds of memory?
| Kind | What it holds | Relation to these two |
|---|---|---|
| Autobiographical memory | your life story | Overlaps both. Your first day at school is an episode, and its name is a personal fact |
| Working memory | what you’re using right now | Short-lived. Both feed it, and its “episodic buffer” is a different thing from episodic memory |
| Procedural memory | how to do things | Usually placed apart from both, which are declarative |
| Sensory memory (iconic, echoic) | the raw sight or sound, for moments | Comes before either of them |
What do people get wrong about them?
That Tulving invented the idea from nothing
Simply Psychology says “No one had drawn this line before”, and that Tulving “first proposed the term ‘semantic memory’”, though its semantic memory page does credit Reiff and Scheerer. Tulving wrote of his own distinction that it “was not at all original”. The term was Quillian’s. Tulving noted similar lines in philosophy, such as Bergson and Russell, and wrote that they hadn’t influenced psychological research, Reiff and Scheerer aside. In 1987 he added that students of memory pathology had said similar things.
That the 1972 distinction was about the feeling
You’ll read that in 1972 he identified remembering as a feeling and knowing as recalling facts. On Renoult and Rugg’s account, the 1972 chapter contrasted what each stores and refers to, and the emphasis on experience was added in 1983 and 1985.
That they’re two separate boxes
Renoult and Rugg say his contribution is “sometimes summarized in terms of a rigid distinction”, though from the first he recognised the two as highly interactive. He did argue from 1983 that they’re functionally distinct. In 2005 he wrote that it isn’t necessary to assume the line between knowing and remembering “is always sharp”. The 2019 review calls the two “inextricably” linked, and the 2020 one says they keep “a degree of distinctiveness”.
That remember and know settle which memory you used
Tulving mapped remember and know onto episodic and semantic memory. Kihlstrom reads Gardiner differently: remember is recollection and know is familiarity, a split within recognising things. On that reading, a “know” in the demo above isn’t necessarily semantic memory.
Asking people about their own experience has limits too. Yonelinas and colleagues report that these answers converge fairly well with more objective measures, and that the instructions “can sometimes be misunderstood”.
That semantic memory is just words
The 1972 definition was tied to language. Tulving moved away from it in 1983, since “we know many things about the world that are neither meaningful, nor readily expressed in words”.
Do AI memory systems use this split?
Yes, as tiers and labels. The CoALA framework for language agents says “Semantic memory stores facts about the world” and episodic memory stores experience from earlier steps, and several systems keep the two in separate tiers. A label doesn’t settle whether they have the part Tulving came to care about most, the sense of going back.
The Field Guide’s page on this mechanism covers which systems do what, including Counterparts, the memory system whose site this is, and the directory rates them side by side.
Questions people ask
What is episodic memory in simple terms?
Your memory for things that happened to you: a particular event, at a particular time and place. Remembering it feels like going back.
What is semantic memory in simple terms?
What you know about the world: facts, the meanings of words, what things are for. You can use it without recalling when you learned it.
What are examples of episodic memory?
Your first day at a job, what you had for dinner last night, the conversation you had on a train. Tulving's sample sentence was about meeting a retired sea captain who knew more jokes than anyone the speaker had met.
What are examples of semantic memory?
That Paris is the capital of France, what a bicycle is for, the meaning of the word "thesaurus", and your own date of birth.
What is the difference between episodic and semantic memory?
Episodic memory is tied to your own past, with a when and where. Semantic memory is knowledge detached from how you got it. A quick test: can you say when you learned it, and does anything come back? They're still closely linked, as Tulving himself said.
Who came up with episodic and semantic memory?
Endel Tulving proposed the distinction in a 1972 chapter. The term semantic memory was already in use, from Quillian in 1966, and Tulving took the word episodic from "nonepisodic memory" in a small book by S. Munsat.
Are episodic and semantic memory part of long-term memory?
Yes. Both count as declarative memory, the kind you can state, and procedural memory is usually placed separately.
Is the episodic buffer the same as episodic memory?
No. The episodic buffer is a part of working memory, a short-lived store added to Baddeley's model in 2000. Episodic memory is long-term. Baddeley's word reflects binding things together into one episode.
Where are they in the brain?
They overlap a lot. Simply Psychology's summary puts the medial temporal lobe, including the hippocampus, at the centre of episodic memory, and Binder and Desai's review finds semantic processing in the inferior parietal lobe and much of the temporal lobe. Renoult and Rugg describe the overlap as extensive.
Do animals have episodic memory?
It's argued about. Scrub jays remember what they cached, where and when, so Clayton and Dickinson called this episodic-like. Their tests measure behaviour, so they can't show the felt sense of going back that Tulving made central.
What should I write in an exam?
Tulving (1972): episodic memory stores personally experienced events with their time and place, and semantic memory is organised knowledge. Then give an example of each, then Tulving's later additions (1983, 1985): the felt sense of remembering, and the remember/know test. Evidence for the split includes brain damage that affects one more than the other, such as the patient K.C. (Rosenbaum and colleagues, 2005). Add if you have room that the two interact constantly, and that the line is less sharp than two boxes suggest.
Questions about how any of this carries over to AI? Ask in r/AIMemorySystems.
References
Sources
- Renoult & Rugg (2020). An historical perspective on Endel Tulving's episodic-semantic distinction. Neuropsychologia. Read in the parts cited, from the accepted manuscript. Our main source for what Tulving's 1972 chapter and 1983 book say.
- Tulving (1987). This Week's Citation Classic: Episodic and semantic memory. Read in full. Tulving's own account of writing the 1972 chapter.
- Kihlstrom (2019). Varieties of recollective experience. Neuropsychologia. Read in the parts cited. Source of the chart, and our source for Gardiner (1988), Tulving (1985) and Conway et al. (1997).
- Yonelinas et al. (2010). Recollection and familiarity: examining controversial assumptions and new directions. Hippocampus. Read in the parts cited.
- Rosenbaum, Köhler, Schacter, Moscovitch, Westmacott, Black, Gao & Tulving (2005). The case of K.C.: contributions of a memory-impaired person to memory theory. Neuropsychologia, 43(7), 989-1021. Abstract only.
- Renoult, Irish, Moscovitch & Rugg (2019). From knowing to remembering: the semantic-episodic distinction. Trends in Cognitive Sciences. Abstract only.
- Schacter & Addis (2007). The cognitive neuroscience of constructive memory. Philosophical Transactions of the Royal Society B. Read in the parts cited.
- Binder & Desai (2011). The neurobiology of semantic memory. Trends in Cognitive Sciences. Read in the parts cited.
- Clayton & Dickinson (1998). Episodic-like memory during cache recovery by scrub jays. Nature. Abstract only.
- Tulving (2002). Episodic memory: from mind to brain. Annual Review of Psychology. Abstract only.
- Sumers, Yao, Narasimhan & Griffiths (2024). Cognitive architectures for language agents. Transactions on Machine Learning Research. Read in the parts cited.
- Simply Psychology, Episodic memory and Semantic memory. A secondary source, used for what summaries say and for the medial temporal lobe.
Last checked on . Tulving’s 1972 chapter, 1983 book and 1985 paper, and Gardiner (1988), weren’t available to us, so the notes above say whose description we relied on. Made by Counterparts AI. Counterparts is an open source memory for AI. If something here is wrong, tell us.