Short-term and long-term memory: what’s the difference?

Try it

Watch the last words of a list disappear#

This is free recall, the task behind much of the classic evidence. Compare what comes back straight away with what comes back after a short distraction.

Straight away, the last few words usually do best, and they tend to be the first ones people write down. The classic account: they’re still in short-term memory, so you can read them out. Counting pushes them out, and everything else has to come from long-term memory.

It’s a demonstration, shorter and faster than the original experiments, with no norms behind it. Your words and scores stay on this page, and we only count that a demo was started.

If you can’t see it: twelve single words appear one every second and a half, marked so a screen reader should announce each. You type back what you remember. On alternate lists you first subtract 3 repeatedly for 20 seconds in a text box. The page then shows which positions you recalled.

How did anyone tell them apart?#

The idea is old. In 1890 William James separated primary memory, what’s still in mind, from “memory proper, or secondary memory”: knowing about something “of which meantime we have not been thinking”. He quotes Exner’s example: absorbed in something, you don’t hear the clock strike, then your attention wakes and “we may then count off the strokes”.

On that test, what you had for breakfast is long-term memory, even though it was this morning. It left your mind and came back.

The evidence from free recall

In the 1960s, free recall gave the idea numbers. The chance of recalling a word depends on where it was in the list, and Murdock (1962) and Glanzer and Cunitz (1966) showed that the end of the list behaves differently from the rest.

Two things that can be changed separately suggest two separate things underneath. The last row is the most striking: the words you were best at a few minutes ago become the ones you’re worst at. A common explanation is that they spent less time in the buffer, so they had less time to be stored.

The model most courses teach

Atkinson and Shiffrin’s 1968 chapter drew this as sensory, short-term and long-term stores, and it’s the usual starting point. Their 1971 paper opens with “Memory has two components: short-term and long-term”.

Studies of people with brain damage are another large part of the classic evidence. This page doesn’t cover them.

The differences, side by side#

That last row uses the words differently. In Eric Kandel’s sea slugs, a single shock to the tail gives a “short-term memory” lasting minutes, and four or five spaced shocks give a memory lasting days. Psychology’s short-term is seconds, so when someone mentions it, check the timescale.

Working memory is a related term and it has its own page, including how much short-term memory holds and how the two differ. Echoic memory, the brief copy of a sound, comes before all of this.

How long does short-term memory last?#

Seconds, and the popular numbers are firmer than the evidence. Peterson and Peterson (1959) gave people three consonants, had them count backward by threes, and tested them after up to 18 seconds. Simply Psychology reports about 70% forgotten at 9 seconds, and about 90% at 18.

If you can’t see it: three letters appear for a second and a half, then you subtract 3 repeatedly in a text box for 3, 9 or 18 seconds, then type the letters. It’s a demonstration, with far fewer trials than the original.

Sceptics argued that the letters got harder to pick out from those on earlier trials as the delay grew. So Baddeley and Scott (1971) set up a trailer in a shopping mall and tested a large number of people once each. They found a drop within the first 5 seconds and nothing beyond it.

Nelson Cowan’s verdict in 2008 was that decay “is not yet on very firm ground”. He also raises the possibility that a memory doesn’t fade gradually, and instead collapses at a moment that varies from trial to trial.

How does something get into long-term memory?#

The classic answer was rehearsal. Waugh and Norman (1965) proposed that it both keeps items in short-term memory and transfers them to a lasting store, and Craik and Watkins say Atkinson and Shiffrin endorsed that in general.

Craik and Watkins tested that in 1973. People heard lists and had to keep hold of the latest word starting with, say, G: GARDEN, then GRAIN, then GIRAFFE. How long a word stayed in mind varied. In a surprise test on all the lists afterward, that time didn’t predict recall.

They concluded: “Maintenance rehearsal does not lead to an improvement in memory performance.” Their account is that elaborative rehearsal, working on what an item means and connecting it to other things, is the kind that matters.

Not everyone puts it so strongly. Malmberg, Raaijmakers and Shiffrin, one of them an author of the 1968 chapter, answer the related critiques by Craik and his colleagues in 2019: they say “modest benefits of maintenance rehearsal and added benefits of elaborative rehearsal” have been confirmed many times, and that the chapter already made the point.

You also don’t need to be trying. Norris points to Hebb’s 1961 result: people repeating lists straight back do better on a list when it comes around again, so long-term learning runs during tasks that look purely short-term.

At the level of cells, Kandel’s sea slug work found that the long-term form needs new protein, and that spaced repetition is what produced it. What happens across a whole brain in the days after learning is on the Field Guide’s consolidation page, and how recalling something strengthens it is on retrieval practice.

How much can long-term memory hold, and for how long?#

Nobody has found the ceiling. In the 1970s, people who saw 10,000 scenes for a few seconds each could later pick the one they’d seen out of a pair 83% of the time. That could have been gist alone, so Brady and colleagues tested for detail in 2008.

Time is the other half. A study of 392 high school graduates, reaching back about 47 years, found that recognising classmates’ names and faces stayed around 90% correct for the first 15 years. Then names fell, while faces held until about 35 years. Recalling names without help declined steadily.

Course material behaves differently. Conway and colleagues tested 373 former students of a cognitive psychology course, up to 12 years on. Names of researchers and concepts fell for about three years, names faster, then held above chance. General facts and research methods stayed at the same above-chance level across the years they sampled.

Long-term memory covers what happened to you and what you know, as well as skills you can’t put into words, such as riding a bike. How it fades is on the Field Guide’s page on the forgetting curve.

Are they really two separate systems?#

Researchers still disagree about whether the differences across seconds and years need two stores. Most models since the 1960s assume a separate store, Norris says, and the view that short-term memory is activated long-term memory has become increasingly popular in neuroimaging research.

The case for one system

Cowan and others hold that short-term storage is the part of long-term memory active right now. Single-store models get a strong argument from recency, which isn’t only short-term. Bjork and Whitten (1974) put a distracting task after every word, including the last, and the last words still did best. On the classic account the buffer should have been cleared, so a single rule might explain recency at every timescale.

Even the founders hedged. Atkinson and Shiffrin wrote in 1971 that “one might consider the short-term store simply as being temporary activation of some portion of the long-term store”, and Shiffrin wrote in 1975 that “STS is the activated subset of LTS”. The 1971 paper still draws separate boxes, and Norris says: “Nothing seems to rest on this assumption.”

The case for two

Norris’s 2017 review argues that activation can’t do the job. You can repeat “blontestaping”, which you’ve never heard, so there was nothing in long-term memory to activate. And activating “1” and “3” can’t represent the sequence 1, 3, 1, where the same digit appears twice in a set order.

Davelaar and colleagues add that the two kinds of recency come apart. Negative recency shows up after immediate recall and is absent in the version with a task after every word, so they argue both a buffer and a changing context are needed. Poltrock and MacLeod (1977) called the results of the every-word version “completely consistent” with multistore accounts.

Cowan and Chen themselves wrote in 2008 that “although the mechanisms of short-term memory are separate from those of long-term memory, they are closely related”. Wickelgren, who thought the balance favoured two stores, complained in 1973 that most of the cited evidence was “worthless for making the distinction”. Norris adds, of the behavioural data: “Very little has changed since then.”

What do people get wrong about it?#

That short-term memory lasts 18 or 30 seconds

Simply Psychology gives “about 15 to 30 seconds” and Wikipedia gives “~18 seconds without rehearsal”. Eighteen seconds was the longest delay Peterson and Peterson tested, as far as we can tell from the descriptions we read. What caused the loss before that point is disputed, as the section above explains.

That Atkinson and Shiffrin’s model only ran one way

Simply Psychology’s multi-store page says “The MSM shows information flowing only one way” and counts that a weakness. The 2019 account, with Shiffrin as an author, says Atkinson and Shiffrin “made it clear that information flowed both ways”. When you read “cow”, what you know about cows joins it in the short-term store. The 1968 chapter itself said “both STS and LTS are active” in experiments on either.

That long-term memory has no limits

Simply Psychology’s multi-store page lists capacity as “Unlimited” and duration as “Unlimited (information in the LTM can potentially last a whole lifetime)”. No experiment could show either, as Simply Psychology’s long-term memory page itself notes. What we have is thousands of pictures remembered, and recognition of classmates that holds for decades and then falls.

Do AI systems have short-term and long-term memory?#

By name, yes. LangGraph’s documentation calls the conversation history “short-term memory, or thread-scoped memory”, and calls data kept “across sessions” long-term memory. The CoALA framework also splits agents into a short-term working memory and long-term stores.

The line there is about scope, and time has little to do with it. A LangGraph thread is saved to a database and “can be resumed at any time”, so its short-term memory can sit unchanged for months. Nothing fades in seconds and nothing needs rehearsing. The limit that does bite is the context window, which the working memory page takes up.

The comparison holds loosely in one place: in both, what’s held now needs something more to last. Here that’s a write to a long-term store, without which one thread’s contents don’t reach the next. Counterparts, the memory system whose site this is, has the AI write down what it learned in its own words.

The Field Guide follows other mechanisms through to what AI systems did with them, and the directory compares the systems.

Questions people ask#

Questions about how any of this carries over to AI? Ask in r/AIMemorySystems.