THE CONTROL PROCESSES OF SHORT-TERM MEMORY
by
R. C. Atkinson and R. M. Shiffrin
TECHNICAL REPORT 173
April 19, 1971
PSYCHOLOGY SERIES
Reproduction in Whole or in Part is Permitted for
any Purpose of the United States Government
This research has been supported by the National Science
Foundation under grant NSFGJ-443X.
INSTITUTE FOR MATHEMATICAL STUDIES IN THE SOCIAL SCIENCES
STANFORD UNIVERSITY
STANFORD, CALIFORNIA
THE CONTROL PROCESSES OF SHORT-TERM MEMORY
R. C. Atkinson and R. M. Shiffrin
Stahford University
Stanford, California 94305
Human memory is divided into a short-term working memory and a long-term permanent memory. Control processes act within the short-term working memory to make decisions and regulate information flow, thereby controlling learning and forgetting.
The system by which information is stored in and retrieved from memory has always been a topic of great interest to psychologists. The English associationists and early experimental psychologists like Wilhelm Wundt, William James, and Ernst Meumann relied upon introspective techniques to generate their theories. Their introspections led them before the turn of the century to divide memory into short-term and long-term components. They discerned a clear difference between thoughts currently present in consciousness and those that could be brought to consciousness after a search of memory that often required considerable effort. For example, this sentence is in your current awareness, but the winner of the 1968 World Series, while probably in memory, requires some effort to retrieve and in fact may not be found at all.
Despite its intuitive attractiveness, the short- versus long-term view of memory was largely discarded when psychology turned to behaviorism which emphasized animal as opposed to human research. The short- versus long-term distinction received little further consideration until the 1950's when a number of psychologists, particularly Donald Broadbent in England, Donald Hebb in Canada, and George Miller in the United States,
---1---
reintroduced it (see George A. Miller, Information and Memory, Scientific American, 1956, 195 (2), 42-47). The growth of two-process systems was accelerated by the concurrent development of computer models of behavior and mathematical psychology. The two-process viewpoint is now undergoing considerable theoretical development and is the Subject of a large research effort. In particular, the short-term memory system, which we will callshort term store(or STS) , has achieved a position of pivotal importance. Its importance stems from processes carried out in STS that are under the immediate control of the Subject. Thesecontrol processesgovern the flow of information in the memory system; they can be called into play at the Subject's discretion, with enormous consequences on performance.
Some control processes are used in many situations by everyone, and others are used only in special circumstances.Rehearsal, an overt or covert repetition of information, is employed in numerous situations: when remembering a phone number until it can be written down, when remember; i.ng the names of a group of people to whom you have just been introduced, and when copying a passage from a book, to name a few examples.Codingrefers to a class of control processes in which long-term retrieval is enhanced by placing the to-be-remembered information in a context of additional and easily retrievable information. For example, students sometimes learn the twelve cranial nerves with the use of the mnemonic "OnOldOlympus'TinyTopAFinnAndGermanViewedSomeHops" where the first letter of each word corresponds to the first letter of each nerve.Imagingrefers to a control process in which verbal information is remembered through the use of visual images. The ancient Greeks made
---2---
extensive use of this process. Cicero suggested learning long lists (or speeches) by placing each member of the list in a visual representation of successive rooms of a well-Known mansion. A number of other controlprocesses, including decision rules, organizational schemes, retrieval strategies and problem solving techniques, will be encountered in this article. The point to Keep in mind is the optional nature of control processes. In contrast to permanent structural components of the memory system, the control processes are selected at the Subject's discretion; they may vary, not only with different tasks but even from one encounter with the same task to the next.
An Outline of the Memory System
We believe that the overall memory system is best described in terms of the flow of information into and out of STS and the Subject's control of this flow. Before describing the system it is helpful to introduce terminology with which to discuss information flow. All phases of memory are assumed to consist of small units of information which are associatively related. Any set of information units that are closely interrelated will be termed an "image" or "trace" (thus "image" does not necessarily imply a visual representation). If the pair "TKM - 4" is presented for memory, the image stored might include the size of the card on which the pair is printed, the type of print, the sound of the various symbols, the semantic codes, and numerous other units of information.
The basic phases of the memory system and the types of information flow are illustrated in Figure 1. Information from the environment is accepted and processed by the sensory registers in the various sensory
---3---
Figure Caption
Figure 1. Information flow in the memory system. Environmental information is processed by sensory registers in the various physical modalities and entered into short-term store (STS). The information remains temporarily in STS, the length of stay depending on control processes. While information remains in STS it may be copied into long-term store (LTS). While information remains in STS, information in LTS associated with it may also be activated and entered in STS. Thus, if a picture of a triangle is presented, this visual information is processed and entered into STS. Then, since the verbal name "triangle" is associated with this visual information in LTS, this verbal label is also entered into STS.
modalities, and entered into STS. Information resides in STS for a period of time that is usually under the control of the Subject. By rehearsing one or more items the Subject can keep them in STS, but the number that can be maintained in this way is strictly limited. For example, most people can maintain 7 to 9 digits. Once an image is lost from STS, it cannot thereafter be recovered from STS. During the period in which information resides in STS it may be copied into long-term ~ (or LTS); we shall see that the transfer of information from STS to LTS is highly dependent upon rehearsals of that information in STS. LTS is assumed to be a relatively permanent memory store, from which information is not lost. Information is copied from LTS to STS as well as in the reverse direction; in fact, we assume that during the period an image resides in STS, some information in LTS closely associated with that image will be activated and also entered into STS. Thus information entering STS from the sensory registers will initially be specific to the modality of input, but almost at once close associations from LTS in all modalities will be activated and placed in STS. For example, a word may be presented visually, but immediately after input the articulatory-verbal "name" and associated meanings will be activated from LTS and placed in STS.
Our account of STS and LTS does not require that the two stores necessarily be in different parts of the brain, or involve different physiological structures. It is possible, for example, to view STS simply as a temporary activation of some portion of LTS. The same physiological structures might be involved in both instances, the only distinction being whether or not a given structure is currently activated. Also, in our thinking we tend to equate STS with "consciousness"; the
---4---
thoughts and information of which we are currently aware can be considered to be part of the current contents of STS. Such a statement lies in the realm of phenomenology, and as stated cannot be scientifically verified. Nevertheless, thinking of STS in this way may help the reader conceptualize the short-term system. Because consciousness is equated with STS, and because control processes are centered in and act through STS, this store is considered to be a "working memory": a store in which decisions are made, problems are solved, and information flow is directed.
Retrieval of information from STS is quite fast and accurate. Experiments by Saul Sternberg at Bell Telephone Laboratories and others have shown that the retrieval time for information in STS Such as letters and numbers ranges from lO to 30 milliseconds per character. The retrieval of information from LTS is considerably more complicated. So much information is contained in LTS that the major problem is finding access to some small subset of this information which contains the desired image. This problem might be likened to the task of locating a particular book in a library. Once the book is located, it may then be scanned in an attempt to recover the desired information. We propose that the Subject activates a likely subset of information, places it in STS, and then scans STS for the desired image (which may not be present in the current Subset). The retrieval process therefore becomes a search in which various Subsets are activated and scanned. The conception is depicted in Figure 2. On the basis of the test query the Subject selects a small set of features termed "probe information" and places the probe in STS. The subset of information in LTS closely associated with the probe will then be activated and entered into STS; this Subset is termed
---5---
Figure Caption
Figure 2.Information flow and decisions during the search of long-term memory. The probe is placed in STS, then information in LTS closely associated with the probe is activated and placed in STS. This set of information is called the "search-set." Before the search-set is lost from STS, the Subject draws images from the search-set for examination. If the desired information is not found, the search is either stopped, or recycled to another selection of probe information.
the "search-set." The Subject selects from the search-set some image, which is then examined. The information extracted from the selected image is utilized for a decision: has the desired information been found? If so, the search is terminated. Even if the information has not been found, termination may occur if the Subject decides continuation is unlikely to be productive. If the search does continue, the Subject begins the next cycle of the search by selecting a probe once again. This may or may not be the same probe used on the preceding cycle, depending upon the Subject's strategy. For example, the Subject may be asked to search for states of the United States starting with the letter M. He may do so by generating states at random and checking their first letter (in which case the same probe information may be used on each search cycle) or he may generate successive states in a regular geographic order (in which case the probe information is systematically changed from one cycle to the next). It can be shown that strategies in which the probe information is systematically changed will more often result in successful retrieval, but will take longer to do so than alternative "random" strategies. Note that the Freudian Concept of repressed memories would be handled in this framework by an inability of the Subject to generate an appropriate probe.
The Effects of Rehearsal
The reader has undoubtedly noticed that this theory portrays the memory system almost entirely in terms of the operations of STS. This is quite intentional. In our view, information storage and retrieval is best described in terms of the flow of information through STS, and in terms of the Subject's control of the flow. One of the most important
---6---
of these control processes is rehearsal; rehearsal is an overt or covert repetition of information that either increases its momentary strength in STS or otherwise delays its loss. Some examples have been mentioned earlier; other uses of rehearsal occur during the taking of lecture notes or .during the act of performing mental arithmetic. Rehearsal can be shown not only to maintain information in STS but also to control transfer from STS to LTS. We will present several experiments concerned with an analysis of the rehearsal process.
The research in question involves a memory paradigm known as "free recall." Because the experiments to be considered here are based on one or another variation of this paradigm, it will be described in some detail. The situation is analogous to one in which you are asked to name the people present at the last large party you attended. The experimental procedure is extremely simple. A list of random items (usually common English words) is presented to the Subject one at a time. Following presentation the Subject attempts to recall as many words as possible in any order. Many psychologists have worked with this paradigm and major research efforts have been carried out by Bennet Murdock at the University of Toronto, Endel Tulving at Yale University, and Murray Glanzer at New York University. The result of principal interest is the probability of recalling each item in a list as a function of its serial presentation position. Plotting this function yields a U-shaped curve of the form presented in Figure 3a. The increased probability of recall for the first few words in the list is called theprimacy effect; the large increase for the last 8 to 12 words is called therecency effect. There is considerable evidence that the recency effect is due to retrieval
---7---
Figure Caption
Figure 3.The probability of recall as a function of serial presentation position for various free recall experiments. In free recall a list of words is sequentially presented to the Subject and then he is asked to recall them in any order. (a) The basic serial position curve: the rise on the right is called the recency effect because these words are the most recently presented; the rise on the left is called the primacy effect because these are the first words presented. (b) If an arithmetic task is interpolated between presentation and recall the recency effect disappears but earlier portions of the curve are unaffected. The recency effect thus appears to be due to retrieval from STS, and is eliminated when arithmetic causes the words to be lost from STS. (c) The list length effect demonstrates LTS retrieval failure in free recall. Words in long lists are recalled less well than words in short lists. (d) The time available to rehearse each word affects LTS retrieval: slower presentation results in better recall. The above graphs are idealized, and are based on experiments reported by James Deese, Bennet Murdock, Leo Postman and Murray Glanzer.
from STS, and that the earlier portions of the serial position curve reflect LTS retrieval only. In one paradigm, for example, the Subject is required to carry out a difficult arithmetic task for 30 seconds immediately following list presentation and then asked to recall. One can assume that the arithmetic task causes the loss of all words in STS so that recall reflects LTS retrieval only. Indeed, the recency effect is eliminated when this experiment is performed; furthermore, the earlier portions of the serial position curve are unaffected (see Figure 3b).
Variables that influence LTS but not STS also can be manipulated.In these cases, the recency portion of the serial position curve should be relatively unaffected, while the earlier portions of the curve should show changes. One variable that affects LTS but not STS is the number of words in the presented list. As seen in Figure 3c, a word in a longer list is less likely to be recalled, but the recency effect is quite unaffected by list length. Similarly, increases in presentation rate decrease the likelihood of recalling words prior to the recency region, but leave the recency effect largely unaffected (see Figure 3d).
In free recall experiments many lists are usually presented in a session. If the Subject is asked at the end of the session to recall all the words presented during the session, we would expect his recall to reflect LTS retrieval only. The probability of recalling words as a function of their serial position within each list can be plotted for end-of-session recall and compared with the serial position curve for recall immediately following presentation. The results of Such an experiment are shown in Figure 4. For the delayed recall curve the primacy
effect remains, but as predicted the recency effect is eliminated. In
---8---
Figure Caption
Figure 4.In addition to a recall test immediately following each listpresentation, the Subject may be asked at the end of an experimentalsession to recall all the words from that session. The delayed testshould reflect LTS retrieval only. This prediction is verified by theserial position curve for the delayed test: the increasing recencyeffect is missing. The data are from Fergus Craik.
---8a---
summary, the recency region appears to reflect retrieval from both STS and LTS, whereas the serial position curve prior to the recency region reflects retrieval from LTS only.
In 1965 at a conference sponsored by the New York Academy of Sciences we put forth a mathematical model explaining these and other effects in terms of a rehearsal process. It was postulated that the Subject was rehearsing a small number of words in STS at all times during presentation, including the item most recently presented. The words still being rehearsed in short-term store when the last list item has been presented were assumed to be output at once (giving rise to the recency effect).
The transfer of information to LTS was assumed to be a function of the amount of rehearsal given each item during list presentation. Since the words presented first in the list do not have to share rehearsal with many other items, they were assumed to receive additional rehearsal. This extra rehearsal was supposed to cause more transfer of information to LTS for the first items (thus giving rise to the primacy effect).
This rehearsal model was given a formal mathematical statement and fit to a wide array of experiments. The model provided an excellent quantitative account of a great many results in free recall, including those discussed in this paper. A more direct confirmation of the model has recently been provided by Dewey Rundus at Stanford University. He carried out free recall experiments in which Subjects rehearsed aloud during list presentation. This overt rehearsal was tape-recorded, and compared with the recall results. The discussion is simplified if the term "rehearsal set" is used to refer to those items overtly rehearsed between successive presentations of words. The number of different words
---9---
contained in the rehearsal set was found to start at 1 following the first word presented and then to rise until the fourth word; from the fourth word on the number of different words in the rehearsal set remained fairly constant (at about 3.3) until the end of the list (see Figure 5a). The Subjects almost always reported the members of the most recent rehearsal set when the list ended and recall began. Some particularly interesting data from these experiments are seen in Figure 5b. The figure superimposes on the serial position curve a curve giving the mean number of total rehearsals for items presented in various serial positions. A close correspondence is evident between number of rehearsals and recall probability for words prior to the recency effect; in the recency region, however, a sharp disparity occurs.
These findings provide considerable support for the assumption that LTS storage is a function of the number of rehearsals, and that the recency effect arises from STS retrieval rather than LTS. The hypothesis that storage is a function of the number of rehearsals can be checked in other ways. For example, the recall probability for a word prior to the recency region was plotted as a function of the number of rehearsals received by that word. The result was an almost linear, sharply increasing function. Furthermore, consider words presented in the middle of the list that happened to be given the same number of rehearsals as the first item presented. The recall probability for Such items was identical to that for the initially presented item.
Having established the efficacy of rehearsal both in storing information in LTS and maintaining information in STS, an experiment was carried out .in which the Subjects' rehearsal was manipulated directly.
---10---
Figure Caption
Figure 5.(a) Example of a Subject's rehearsal protocol. A partiallisting of a Subject's overt rehearsa19 ir) an experiment by Dewey Rundus.Note that the first word presented ~s given more total rehearsals thanlater words. (b) Probability of recall compared with total number ofrehearsals, at each presentation position. Prior to the recency region,rehearsals and recall are closely related. Thus, LTS storage dependson the number of rehearsals given an item, and the storage differencesappear in LTS retrieval.
---l0a---