What is echoic memory?

Try it

Can you hear the noise repeat?#

Some rounds are a slice of static on a loop: a quarter of a second, one second or four. Others are fresh static that never comes back.

Two passes of static have nothing in common except the sound itself, and it has no words or tune you could name or hum. So if you notice a repeat, some memory of the waveform carried over from one pass to the next. That’s the logic of the task Guttman and Julesz introduced in 1963.

It shows that kind of memory at work without proving it’s the echo: listeners can also learn a particular noise with repetition, and a four-second loop may draw on more. It’s a demonstration, with too few rounds to measure anything, and your headphones, volume and room all change the result. What you hear and answer stays on this page, and we only count that a demo was started.

If you can’t hear it: each round is a steady hiss. In a repeating round, an identical stretch of hiss comes back on a fixed beat; in the others it never comes back. The task can’t be done without hearing. Suffix experiments like the one below have been run in American Sign Language, though, so the broader idea reaches past the ears.

How did anyone find this?#

The name came first. Neisser described echoic memory in 1967 as the ability to “hear” a sound for some time after it has stopped, by analogy with a store he’d named for vision. In 1960 George Sperling had flashed arrays of characters and shown that, if a cue came quickly, people could report most of any row they were cued to. If the cue came about a second late, most of that was gone. That store is now called iconic memory, and its page goes through his numbers.

Two lines of experiment then did similar work for hearing.

The three-channel experiment

Darwin, Turvey and Crowder (1972) had listeners hear spoken lists that seemed to come from three places, then report either everything or just the part a cue pointed to.

• When the cue named a location, the part beat the whole, so long as the cue came less than 4 seconds after the sounds ended.
• When it named a category (letters or digits), the gain was much smaller and held only under 2 seconds.
• The authors summed this up as “a store which has a useful life of around two seconds”.

The “zero” experiment

People recall the last item of a spoken list unusually well. Crowder and Morton (1969) played 36 undergraduates lists of 10 to 30 digits and asked them to write as many of the final digits as they could. On some lists the word “zero” came at the end, and listeners knew it wasn’t part of the list.

• That one extra word hurt recall late in the list and largely left the earlier positions alone.
• It hurt more when digits came faster. At four digits a second the effect was largest, and at one digit every two seconds it was within experimental error.
• In a second experiment, 24 people saw nine digits projected one at a time, with a spoken “zero” half a second after the last. There was no such effect on the last digits.

Crowder and Morton’s reading: an acoustic store holds the end of a spoken list, it decays, and the “zero” displaces what’s left. Later work added that a visual or non-speech suffix does far less damage than speech, and that the suffix hurts more the more it sounds like the last item, whatever it means.

An illustration, not a study: someone reads you a phone number and the last digits feel safe, until they add “okay?” straight after.

Try the suffix task

If you can’t hear it: a voice reads eight digits, about one every three-quarters of a second, then either says “zero” or plays a short tone. You type the eight digits back, and the totals compare the last digit and the first four across the two endings. It’s a demonstration: our lists differ from Crowder and Morton’s, and eight of them can’t measure anything.

How long does echoic memory last?#

No single number holds up. You’ll see “3 to 4 seconds” and “2 to 4 seconds”, which sit close to Darwin’s cut-offs, though the authors’ own summary was “around two seconds”. Nees’s 2016 review lists published estimates from 2 seconds or less up to possibly 60.

Part of the gap comes from what each study counts as an echo.

• Partial report asks you to say what you heard, so the echo has to be read out before it’s gone.
• Comparing sounds asks whether two sounds match. Sams and colleagues tested 13 young adults on tones, who beat chance at gaps of 0.75, 1.5, 3 and 9 seconds but not at 6 or 12, for an estimate of about 9 (the paper’s title rounds it to 10). Their brain recordings gave about 6 seconds on one measure, which they call too short to explain the behaviour, and 9 to 12 or longer on the others.
• Brain responses use the mismatch negativity, a response to a sound that breaks a pattern. A 2020 review puts its window at 7 to 20 seconds.
• Recall after long gaps gives the longest numbers. A spoken list’s advantage over a seen one survived 60 empty seconds in Engle and Roberts’s study, and a suffix delayed by 20 seconds still hurt recall of the last item.

There may be more than one thing being measured. Nelson Cowan’s 1984 review proposed a short store of a fraction of a second and a longer one lasting seconds. In Balota and Duchek’s 1986 study, a suffix delayed by 20 seconds still hurt recall, but the speaker’s voice mattered only if the suffix came straight away. Winkler and Cowan read that as fragile specific detail plus a more durable broad pattern.

What is it for?#

A sensory store gives you a short time to get at information before it’s lost. The studies here point to a few uses.

• Going back for what you missed. Sams and colleagues open their paper with this case: someone reading the paper hears “What do you think of it, dear?” and can “go back in time” to answer.
• Noticing that something changed. Inui and colleagues point out that detecting a change needs a memory of what came before.
• Telling near-identical sounds apart. People can generally only distinguish two very similar sounds if they come within about 10 seconds of each other, according to Winkler and Cowan’s review.
• Holding speech sounds, unevenly. Detail about stop consonants like “b” and “d” fades much faster than detail about vowels. Modality and suffix effects also appear for syllables that differ in a vowel and vanish for ones that differ only in a stop consonant.

How is it different from iconic memory and working memory?#

Echoic memory is easily confused with the phonological loop, the part of Baddeley and Hitch’s working memory model that holds speech. Both deal with sound, and both are short: the loop’s traces were first estimated to fade in about 2 seconds too, a figure later questioned.

The difference is what keeps them going. You keep the loop’s contents alive by saying them silently to yourself. Crowder and Morton’s store, by their description, is “not affected by silent rehearsal”.

What do people get wrong about it?#

That it’s a settled, separate box

Crowder and Morton’s 1969 store held the last item or two for a second or two, as Penney describes it. Penney’s 1989 review set out the problems. The auditory advantage can stretch over 4 to 9 items, it shows up after 20 seconds and more, and it’s counted in words whatever their length.

Nairne’s 1990 model treats the echo as one part of a single memory trace. A 2017 paper says the later evidence “clearly questions” whether the effect comes from a store that is precategorical, acoustic or a store.

That its capacity is known

Some summaries call it high capacity. The store the suffix effect was built on held one or two items, by Penney’s account. Darwin’s partial-report advantage is usually taken to mean more is held than can be reported. The sources here don’t settle it.

That it only comes from the ears

Crowder and Morton said the store “receives information only from the ears”. Suffix effects were then found for lip-read words, American Sign Language and touch. In one study the same physical sound had a different effect depending on whether listeners were led to believe it was speech, which is awkward for a store said to come before meaning.

That memory for sound is gone in seconds

People remember voices long after, and listeners who kept hearing particular random noises remembered them for several weeks.

That mismatch negativity is a direct readout

It’s often used as an index of echoic memory. A 2020 review calls the dispute ongoing: it may reflect a comparison with a memory trace, or neurons tuned to a repeated sound simply responding less. Results also differ on whether attention changes it.

Is there anything like this in AI memory?#

Not really. A sensory store holds a sound for the moment between hearing it and attending to it. A text assistant gets words that are already words, and the next sentence doesn’t overwrite the last.

The nearest thing is a smaller lesson: keep the raw input for a while, and decide what it meant later. On a local install, Counterparts, the memory system whose site this is, keeps conversation text only until it’s written up, deletes it 3 days after that, or about two weeks after the conversation goes quiet. That runs on days, on disk, by design.

The Field Guide covers mechanisms AI systems have tried to copy, such as forgetting and consolidation.

Questions people ask#

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