What is iconic memory?

Try it

Do you see more than you can say?#

This is partial report, Sperling’s method, on your own screen. You go through twelve rounds, and at the end it sets your numbers beside his.

The trick is that you can’t know which row is coming. If you get three of four in a row you were only told about after the letters had gone, you must have had about three of four in every row, which is nine of twelve. The box is black in both light and dark themes because Sperling’s fields before and after the flash were dark, and a bright field afterward cut his observers’ scores to about half.

It’s a rough version: arrows where he used tones, far fewer trials, and no practice, where his observers worked for weeks. Your letters and scores stay on this page, and we only count that a demo was started. A screen can’t match a tachistoscope, so the demo times itself and tells you what it managed. A letter counts only in its own box, as in Sperling’s tables: he scored it both ways, found his procedures didn’t affect one score apart from the other, and tabulated right letter in the right place because it’s less open to guessing.

If you can’t see it: this demo has no version that works without sight, since the effect is about vision. In place of the flash, the table in the next section gives Sperling’s results.

How did Sperling find it?#

He began with a known puzzle: observers, in his words, “commonly assert that they can see more than they can report”. Flash an array of letters and people could list about four, and that barely moved. Experiments 1 and 2 got an average of about 4.3 with different arrays and little change for exposures anywhere from 15 to 500 milliseconds.

So was four all they’d taken in, or was the limit in getting it out? Asking for everything couldn’t tell those apart, because the report itself might be what ran out.

The partial report

For the arrays, Sperling used three rows of four, flashed for 50 milliseconds on a tachistoscope. Just after the flash a tone sounded: high for the top row, middle for the middle, low for the bottom. Observers wrote only that row. He scored each by what was right on the cued row, multiplied by the number of rows.

• Asked for everything, observers averaged 4.3 of 12, from 3.9 to 4.7.
• Cued as the flash ended, they had an estimated 9.1 of 12 available, from 8.1 to 11.0 (five observers; the chart below has four).
• He called that “only a lower bound”, since partial report can’t give an upper one, and said that with more complex arrays the figure would keep rising.

What happened when the tone came later

He delayed the tone. If the extra letters were sitting in a fading image, the score should fall as the tone came later, and it did.

His summary: for “a few tenths of a second” observers have “two or three times as much information available as they can later report”, and “within one second after the exposure the available information no longer exceeds the memory span”.

Position worked and category didn’t

He also tried cueing by kind: report only the letters, or only the numbers. That gave little advantage over whole report, so position worked as a cue and category didn’t. Chow lists “precategorical”, held before it’s sorted into kinds, among the things iconic memory has been identified with. Wikipedia’s summary instead calls informational persistence post-categorical, and we haven’t traced who holds that.

Who was tested

Five people: four through the student employment service and one faculty member who was interested in the research, in sessions three times a week. That’s small. Sperling lists three investigations with similar techniques, one his own with a different tachistoscope and light source, and says of that one: “All the main findings were reproducible.”

Averbach and Coriell published a related method in 1961, pointing to one position in a row. Neisser named the store in 1967. Sperling’s own phrase was visual information storage.

Is it the same as an image lingering?#

Maybe not. Simply Psychology’s examples run the two together, though the same page separates them. Loftus and Irwin separate two facts, following Coltheart’s 1980 review. “Something that looks like the physical stimulus continues to be present” after it has gone, which is visible persistence. And information can still be picked up from it after it’s gone, which is informational persistence, and that’s what partial report measures.

Coltheart’s case was that longer and brighter stimuli shorten visible persistence and do little to informational persistence, so they can’t be one thing. Later work has questioned how clean that is. Loftus and Irwin’s 1998 experiments then compared measures against ratings of how complete an image looked. Those ratings and temporal integration, where two brief patterns are combined, correlated around 0.98, “something very close to the same thing”. The ratings and partial report correlated only 0.63 to 0.81, so they’re related and not the same.

Where in the visual system is it?

That’s argued about as well. Sligte, Scholte and Lamme (2008) found iconic memory “virtually absent” at equal brightness, which leaves only a weak afterimage, or when flashes of light intervened, and suggest it’s driven by prolonged activity in the retina. Their study also found a second, fragile store after it, lasting about four seconds and holding at least twice as much as ordinary visual short-term memory; a new pattern in the same place overwrites that store.

Öğmen and Herzog (2016) point to a different problem. A store tied to positions on the retina would smear anything that moves, which the visual system then has to control. So they add a sensory store that follows objects in motion.

Is it a useful idea at all?

Chow’s history is blunt: its importance “is no longer recognized by Neisser (1976)”, the man who named it, and Haber (1983) seriously questioned its usefulness, with rejoinders from Loftus and from Chow. Holding’s complaint that the partial-report advantage came from artifacts such as guessing has, according to Chow, been dealt with by Coltheart. We haven’t read those papers, only this account of them.

How is it different from echoic and short-term memory?#

Echoic memory is the same idea for sound, and Neisser named it by analogy. Its page has two listening demos and goes through why its length is less settled. The whole-report limit of about four belongs to the next stage: see short-term and long-term memory and working memory.

What do people get wrong about it?#

That Sperling’s advantage was gone within a quarter to half a second

Simply Psychology says of his experiment, “The advantage faded within a quarter to half a second.” In Sperling’s twelve-symbol average, partial report was still at about 6 letters at 0.3 seconds, well over the 4.3 of whole report, and he puts the point where it meets the span at about 0.5 to 1.0 seconds. A quarter of a second does appear in his paper, for one observer on eight-letter arrays: most letters over that observer’s span gone “within about 0.25 sec”.

The number also depends on the conditions. His fields were dark, and a bright field after the flash lowered scores, which he put down to fewer letters being available or for less time.

That a lingering image is simply iconic memory

It depends on the definition. Simply Psychology says of a room lingering after the light goes off, “That fading impression is iconic memory”, and the same page says researchers treat visible and informational persistence as distinct. On that split, the room is visible persistence, and partial report measures the other one. The two don’t come apart cleanly in the history. Sperling’s own hypothesis was “a rapidly fading, visual image”, and Chow says iconic memory has been identified with a kind of visible persistence. Lightning, in our opening, is only a comparison.

That Sperling found its full capacity

Simply Psychology says partial report “recovers the store’s true capacity”. He didn’t claim that. Nine of twelve was a lower bound, and he thought it “probable” that the capacity he saw “was limited by the small amount of information in the stimuli rather than by a capacity of the observers”.

Do AI systems have anything like iconic memory?#

Not really. A model given an image receives it as input all at once and takes it in as tokens, which Anthropic’s documentation counts in 28-by-28-pixel patches. Those tokens stay available for as long as they stay in the context, and nothing fades in a fraction of a second.

Counterparts, the memory system whose site this is, keeps what the AI writes down in its own words, so what it stores is a description of what happened. The Field Guide follows mechanisms that AI systems do borrow, and the directory compares the systems.

Questions people ask#

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