Primary Memory and Working Memory

The computer model represented an enormous change in the way we thought about memory. It completely refocused memory research. Instead of looking at retention by an undifferentiated memory system, cognitive psychologists began to examine the properties of different storage structures. For a while, most of this research was focused on the sensory registers and primary memory.

The Icon and the Echo

At the lowest end of the multistore model of memory lie the sensory registers -- one for each sensory modality: the icon, the echo, and other analogous sensory memories, in which information arriving from the sensory surfaces is held for a brief time before being copied to primary memory.

In theory, the sensory registers have unlimited capacity, holding all the information arriving from the sensory surfaces.
Information in the sensory registers is represented in pre-categorical form, not yet processed for meaning.
The contents in the sensory registers are retained only very briefly, perhaps for less than a second.
Information in the sensory registers cannot be maintained by any kind of cognitive activity.
Unless they are copied to primary memory, the contents of the sensory registers are permanently lost, either through decay or through displacement by other information newly arriving from the sensory registers.

We know this from experiments employing a paradigm devised by George Sperling (1960), in which subjects were briefly presented (for a few hundred milliseconds) with a 3x4 visual array of letters. After the array disappeared, there was a retention interval of only 0-5 seconds, and then the subjects were asked to report the contents of the array.

In the whole report condition, the subjects were asked to report the entire array: on average, they were only able to report relatively few of the items.
In the partial report condition, the subjects were cued by a tone of low, medium, or high pitch to report the contents of only the bottom, middle, or top rows. Under those conditions, they performed considerably better.
However, even in the partial report condition memory dropped off rapidly over the retention interval.

The implication was that the entire array was actually represented in iconic memory, because the subjects could report accurately the contents of any randomly selected row -- indicating that the contents of all the rows were available. However, in the whole report condition, subjects could not extract all the information in the array before it rapidly decayed.

A similar experiment was performed in the auditory domain -- the echo -- by Darwin, Turvey, and Crowder (1972). In this case, the subjects were presented with 3 lists of 3 items each, for a total of 9 items, presented over stereo headphones in such a way that each list was perceived as coming from a different part of space -- left, middle, and right. Performance in whole-report and partial-report conditions, with the partial reports cued visually, paralleled those obtained by Sperling.

Apparently, the function of the sensory registers is to hold briefly presented stimulus information long enough to allow attention to be directed to it. In this way, the sensory registers exemplify an important function of memory, which is to free behavior from control by whatever stimulus happens to be present in the environment (because, in this case, the stimulus is no longer present).

This is all very well and good, and the Sperling experiment stimulated several years' worth of research activity on the part of a number of investigators. But it's not clear how useful the sensory registers are. In the real world, stimuli usually remain present in the environment for longer than just a few hundred milliseconds -- and this ecological fact may obviate any need for the sensory registers. As Ralph Haber once put it, in a famous paper, iconic memory may only be useful for reading at night in a lightning storm!

Still the sensory registers are memory storage structures -- the point where sensory information first makes contact with the cognitive system. And they do give an organism the opportunity to react to very brief events.

The icon was probably useful in the evolutionary scheme of things, because it enables organisms to catch the movements of predators and prey "out of the corner of the eye", as it were.
The echo, though, was probably even more useful, because many sounds are normally brief -- in particular, the sounds that make up the phonemes of spoken language.

Secondary Memory

At the highest end of the multistore model is a long-term store or long-term memory, also known as secondary memory -- a term that, as noted, was adopted from William James. In theory, secondary memory is the permanent repository of all stored knowledge, both declarative and procedural. For our purposes, these three terms are interchangeable. And, also in theory, the representations in secondary memory reflect considerable processing and transformation. Atkinson and Shiffrin proposed that the capacity of secondary memory was essentially unlimited. Information might be lost from the sensory registers or from primary memory through decay or displacement, but there is essentially no forgetting from secondary memory. Once encoded, memories remain available in storage, and "forgetting" is a matter of gaining access through retrieval processes.

Primary Memory

Primary memory, as conceived in the multistore model, is a short-term store located between the sensory registers and secondary memory. For our purposes, we may use the terms primary memory, short-term memory, and short-term store interchangeably.

In the multistore model, information in the sensory registers is not simply subject to attentional selection it is subject to some actual information processing as well.

Feature detection: first, the stimulus is analyzed into its elementary features.
Pattern recognition: then, some patterns of features are recognized as meaningful.

Here is where the dotted arrow in A&S's diagram comes in. In the theory, knowledge of meaningful patterns is stored in secondary memory. But pattern recognition requires direct contact between the sensory registers and secondary memory. In the Atkinson & Shiffrin model, both feature detection and pattern recognition are unconscious -- precisely because they skirt primary memory, which is the (implied) locus of consciousness in the modal model.

In any event, information in primary memory is subject to further, more extensive processing, as it makes further contact with information stored in, and retrieved from, secondary memory.

Representation in Primary Memory

In the original theory, information was represented in memory in acoustic form -- as, for example, the pronunciation of the letters, numbers, and words in a display or list studied by the subject; or, put another way, as the name of an item, or a description of what the name sounds like.

Capacity of Primary Memory (and Chunking)

Primary memory has limited capacity, as reflected in George Miller's famous "magical number 7, plus or minus 2". That is, primary memory can hold about 7 items of information.

There's a story here. Miller made the capacity of short-term memory famous, but -- as he acknowledged -- it wasn't his discovery. Credit for that goes to John E. Kaplan (1918-2013), a psychologist and human-factors researcher who spent his career at Bell Laboratories, the research arm of AT&T. Prior to the 1940s, telephone exchanges typically began with the first two letters of a familiar word, followed by 5 digits -- as in the Glenn Miller tune, "Pennsylvania 6-500" (the actual telephone number of the Pennsylvania Hotel, often claimed as the oldest telephone number in New York City), or the film Butterfield 8 (after an exchange serving the upscale Upper West Side of Manhattan. But the telephone company (there was only one at the time!) worried about running out of phone numbers, and so switched to an all-digit system. This raised the question of how many digits people could hold in memory. Kaplan did the study, and determined that the optimal number was 7. (Kaplan also determined the optimal layout for the letters and numbers on the rotary, and then the touch pad telephone -- reversing the conventional placement on a hand calculator or computer keyboard -- but that's another story).

However, Miller noted that the effective capacity of primary memory can be increased by chunking, or grouping related items together. Thus, the 18-item string

LBJIRTUSALSDFBICIA

would, ordinarily, be too long to keep in primary memory. But if chunked as follows, the whole thing fits very nicely:

LBJ IRT USA LSD FBI CIA.

So, the capacity of primary memory is roughly 7 chunks, not 7 items. If the chunks are big enough -- yes, you can chunk chunks -- primary memory can hold an immense amount of information.

Of course, chunking in primary memory depends a lot on knowledge stored in secondary memory. You have to recognize a string like LBJ as meaningful -- in fact, the initials of an American president -- before you can chunk it (IRT is the acronym for a subway line in New York City). In fact, the whole 18-letter string is the chorus of a song, "Initials", from HAiR: The American Tribal Love-Rock Musical (lyrics by James Rado & Gerome Ragni): once you've heard the song, you never forget the letters.

The Serial-Position Curve

The distinction between primary and secondary memory is supported by the serial-position curve, which plots the probability of recalling any individual item as a function of its position in the study list. The serial-position curve is typically bowed, meaning that items at the beginning and the end of the list have a higher probability of recall than items in the middle.

The superior recall of items at the beginning of the list is called the primacy effect.
The superior recall of items at the end of the list is called the recency effect.

A common interpretation of the primacy and recency effects is that they reflect the operations of two different memory stores:

The primacy effect reflects retrieval from secondary memory.
The recency effect reflects retrieval from primary memory.

This interpretation is supported by two lines of experimental evidence:

Slowing the pace of list presentation increases the primacy effect, but has no effect on recency. Apparently, increasing the spacing between items gives the subject more time to rehearse each item, and more opportunity to recode items into secondary memory.

Increasing the retention interval, and filling the retention interval with distracting material, reduces the recency effect but has no effect on primacy. The distraction effectively displaces items from primary memory, but has no effect on items already encoded into secondary memory.

The distinction between primary and secondary memory is also supported by neuropsychological evidence from brain-damaged patients.

Wickelgren (1968) noted that the famous patient H.M., who became amnesic following surgical removal of his hippocampus and other structures of the medial temporal lobe, has a normal digit span, and thus normal primary memory. But after a few moments' distraction, cannot remember what he was just doing: he has impaired secondary memory.
Shallice and Warrington (1970) described another patient, K.F., who had a lesion in the left hemisphere at the boundary between the parietal and occipital lobes. This patient had extremely poor digit span -- in fact, he can repeat only a single digit. But he has unimpaired recall of 10-item lists, and thus fairly normal secondary memory.

Both the experimental and clinical evidence indicates that primary and secondary memory can be dissociated:

the operations of the two stores are affected differently by various experimental variables;
the operations of the two stores are affected differently by various forms of brain damage.

However, the strict separation of primary and secondary memory can also be challenged on empirical grounds. For example, primary memory is sometimes referred to as immediate memory. The idea is that items in primary memory are already in consciousness, and immediately available for various operations. By contrast, items in secondary memory appear to be in a latent state: they must be retrieved and brought into consciousness. But it turns out that items in primary memory also need to be retrieved.

This was shown clearly by a classic memory-scanning experiment by Sternberg (1960). In this experiment, subjects were presented with sets of 1 to 6 digits, which they were asked to hold in primary memory (note that 6 digits is well within the normal memory span). Then, they were presented with a probe digit, and asked to say whether the probe was in the memorized list. This is of course an extremely simple tasks, and subjects rarely make any errors in it. But it turns out that the search process takes time. the more items there are in the memory set, the longer it takes subjects to say "yes" or "no". Thus, items in primary memory are not immediately available after all. Primary memory must be searched, just like secondary memory.

As another example, it turns out that there are serial-position curves in retrieval from secondary memory.

American subjects are more likely to remember the names of American presidents like Washington and Jefferson, and Clinton and Bush, than Cleveland or Harrison.
When asked to verify whether letters are part of the English alphabet, subjects are faster with letters like A and B, or Y or Z, than they are with letters like L or M.

Thus, primacy and recency effects do not necessarily reflect retrieval from different memory stores.

The Decline of the Modal Model

What Murdock (1967) called "the modal model of memory" -- the division of memory into three classes of storage structures, and a set of control processes which moved information between them -- had enormous heuristic power. It represented a huge change in the way we thought about memory -- it both exemplified and consolidated the cognitive revolution in the study of memory; and it generated a large number of clever experiments intended to explore the properties of the various storage structures (especially the icon, the echo, and primary memory) and control processes (especially attention). But it also very quickly began to decline.

As noted earlier, the sensory registers are hardly memories at all.

And then there were the doubts raised about the structural distinction between short-term primary memory and long-term secondary memory. Some of these had been raised by Melton (1963), a leading figure in the verbal-learning tradition, even before the modal model had been clearly articulated. But other doubts quickly followed (Murdock, 1972; Craik & Lockhart, 1973; Wickelgren, 1974).

The Paradox of Chunking

The case against a structural distinction was put forth most strongly by Craik and Lockhart (1972), in their analysis of the capacity of primary memory. Certainly primary memory seems to be capacity limited, s in Miller's "magical number 7". But the precise nature of this limitation is unclear.

Is this a limitation on storage capacity, meaning the amount of information primary memory can hold?
Or is it a limitation on processing capacity, meaning the rate at which information held in primary memory can be processed?

And then there is what they called the paradox of chunking. Chunking seems to increase the capacity of primary memory, but chunking also relies on knowledge in semantic memory -- it goes beyond a mere acoustic representation, and includes the results of meaning analysis -- like the meaning of trigrams like LBJ and IRT. This begs the question of coding differences between primary and secondary memory. Primary memory is supposed to represent information in acoustic form -- or, at least, some kind of physical code. But with chunking, the information is clearly not in a purely acoustic (or other physical) form any longer.

The Forgetting Function

And then there is the function of forgetting from primary memory, famously mapped out by Brown (1958) and Peterson and Peterson (1959) in what has come to be known as the Brown-Peterson paradigm. These investigators presented subjects with a consonant trigram such as BNX. Then, they asked the subject to count backwards by 3s from a three-digit number, such as 417, for an interval of 3 to 18 seconds, and then to report the trigram. They observed extremely rapid forgetting: Without distraction, virtually everyone remembered the three letters (though, you do have to wonder about those 5% who didn't!). But , after only 18 seconds of distraction, retention dropped to less than 20% -- a figure that Brown and Peterson attributed to retrieval from secondary memory.

The Brown-Peterson function is consistent with the rapid forgetting claimed for primary memory, but there's a problem. To see what the problem is, try the Brown-Peterson experiment on a friend. Give him a consonant trigram, such as BNX, and have him count backwards by 3s from the number 417 for 18 seconds, and then ask him to write down the trigram. He almost certainly got it right. And if you try it out on your entire dormitory, you might get the occasional instance of forgetting, but nothing like the 80%+ observed by Brown and Peterson.

But that's not exactly how Brown and Peterson did their experiments. Instead of having only a single trial per subject, and averaging over subjects, they gave each subject many trials, and averaged across trials, and then across subjects. Apparently, the Brown-Peterson function is observed only after subjects have gone through many trials.

To check this, Keppel and Underwood (1962) analyzed the Brown-Peterson function on individual trials, with retention tested after 3, 9, and 18 seconds. On the first trial, there was essentially no forgetting, even after 18 seconds of distraction. On later trials, however, they observed a progressive increase in forgetting. They concluded that the Brown-Peterson function represents the build-up of proactive inhibition.

Apparently, items from earlier trials were apparently interfering with items on later trials. But those items from the early trials are no longer in primary memory -- it doesn't have enough capacity. Thus, the Brown-Peterson experiment, which seemed to confirm rapid forgetting from primary memory, isn't really about primary memory after all, because the items that cause the forgetting aren't being held there.

Interestingly, we can also show release from proactive inhibition by changing the category of the item being memorized. Wickens (1972) gave subjects a series of three trials, with all items drawn from the same category -- e.g., 3 consonants. In this phase of the experiment, he observed a progressive buildup of PI, just as in Keppel and Underwood (1962). On the 4th trial, Wickens shifted the item type for half the subjects -- e.g., from 3 consonants to 3 digits. In this condition, PI disappeared, and recall returned to the level observed on Trial 1. As the 5th and 6th trials continued with the new item type, PI continued to build up again -- again, just as in Keppel and Underwood.

Release from PI shows that the interference effect is controlled by the categorization of the item. But these categories are a matter of secondary memory, which go way beyond any acoustic representation. Forgetting in the Brown-Peterson paradigm is not a function of primary memory; rather, the buildup and release of PI must reflect interference from secondary memory. This phenomenon provides further evidence that the forgetting function determined in the Brown-Peterson paradigm doesn't bear on the nature of primary memory, separate from secondary memory.

The general thrust of these kinds of experiments is to cast doubt on the structural distinction between primary (short-term) and secondary (long-term) memory -- which is the sine qua non of the multistore model.

Working Memory

Still, there is something special about primary memory, which is that we can deliberately maintain items in an active state by means of rehearsal. Baddeley and Hitch (1974) proposed that we rename this memory structure working memory -- because this is where we perform cognitive activities. According to Baddeley and Hitch, working memory consists of a central executive which controls several slave systems:

The phonological loop holds speech-based information, and generates "sub-vocal" speech (i.e., mental repetition) during rehearsal.
The visuo-spatial sketchpad rehearses visual information in the form of mental images.
There are, presumably, other slave systems for other modalities.

The slave systems maintain items in an active state. These are very transient memories: once rehearsal stops, the items disappear from working memory; and newly incoming information will replace old information.

So what's the difference between working memory and primary memory? While primary memory was considered as an intermediary between the sensory registers and secondary memory, this is not the case for working memory. Baddeley and Hitch propose that information is processed directly into secondary memory, with no need for an intermediate way-station like primary memory. Then, when the cognitive system wants to use stored information, it copies it into working memory.

John Anderson (1983) has proposed an expanded conception of working memory, which he considers to consist of those representations -- whether acquired through perception or retrieved from memory -- that is currently in a high state of activation. Other representations are also part of working memory:

current processing goals;
the current environment;
and, at least by implication, the self as the agent of some activity or the experiencer of some event.

Baddeley and Hitch's conception of working memory has become extremely popular among cognitive psychologists, and has stimulated cognitive neuroscientists to search for the neural substrates of working memory in the prefrontal cortex.

This page last modified 05/27/2014.