Encoding refers to the content or form in which information is stored in memory; forgetting is loss of the stored information with the passage of time or with exposure to interfering materials; and retrieval refers to accessing information from memory. Any observation of memory reflects all three components, but measurements of recognition time, reflecting memory search, emphasize retrieval. Recognition time is the time required to respond whether a visually presented test item was part of a previously studied list. For an example of a recognition test following a short-term memory list, see Figure 1. Recognition time and accuracy for individual items in both short-term and long-term memory are consistent with retrieval mechanisms involving parallel, or direct access, operations rather than unguided search through many memories. Some aspects of the increase in recognition time with the length of small short-term memory lists suggest a serial, or sequential, search of the list. However, when more detailed analysis of response times and accuracies is carried out, serial search is ruled out as the sole mechanism of retrieval of the prior experience of an item. Instead, response times inversely related to recency of items in particular list positions. Recency determines the efficiency of a simple, direct retrieval process. In contrast with the direct-access retrieval of item information, information about order generally reflects a slower, serial process of recovery from memory in which list items are searched in order.
Retrieval in Short-Term and Long-Term Memory
Primary, or short-term, memory (STM) for a few recently active concepts is distinct from secondary, or long-term, memory (LTM), which is the repository for all the varied information we retain over extended periods of time (James, 1890, 1950). Short-term memory has been described by some theorists as an active subset of long-term memory (Cowan, 1995). Retrieval from LTM during recognition (of a presented item) has many properties of direct access or content addressability (Gillund and Shiffrin, 1984; Murdock, 1971). Direct-access retrieval occurs when a cue or set of cues makes contact with memory in a unitary process without recourse to a sequence of searches through irrelevant memories. In contrast, recall or production (of a previously experienced but unpresented item) may involve a series of retrieval operations in which memory is successively sampled, using general list cues and previously retrieved items as retrieval cues (Raaijmakers and Shiffirin, 1981).
In the 1960s, strong claims were made that access to STM was fundamentally different from the direct-access retrieval from LTM. STM retrieval was asserted to involve a series of comparisons to all currently active concepts (Sternberg, 1966). Although this view fails to give a full account of memory search under all circumstances, it is a useful starting point.
Recognition Time and STM
Anecdotal information about human memory largely reflects memory failures. Scientists use both memory failures (errors) and measures of retrieval time to infer properties of memory. The ease of memory retrieval is often measured in terms of the average time to recognize (respond to) a test item (response time, RT). In S. Sternberg's (1966) classic short-term memory experiments, subjects decided whether a displayed digit (or, in other studies, letter, word, syllable, or outline shape) had appeared on a short study list of one to six items. One key-press indicated a yes response, another no (see Figure 1a). RT increased approximately linearly for each additional list item (see Figure 1b): mean RT = bR+ sL, (R = Y or N), where L is the list length, s is the additional average RT per list item, and bR is a base time. This characteristic linear (or approximately linear) relationship holds for lists of up to six to nine items (Burrows and Okada, 1975), at which point the increment for additional items decreases sharply, possibly reflecting a transition from list lengths within the capacity of STM to those exceeding that capacity.
Underlying Retrieval Mechanisms
The linear relation of mean RT to list length, where both positive (yes) and negative (no) test items frequently yield approximately the same slope, was interpreted by Sternberg as evidence for a serial and exhaustive set of comparisons between the test item and all items in the list representation in short term memory (see Figure 2a). The slope s of the RT function (the added time for each additional list member) was identified as the time to compare the test item with each item on the list, one at a time, in series. The intercept bR then reflects all the processing mechanisms (encoding of the test stimulus, organization and execution of the response) that do not depend on list length. If the list items were searched in series, equality of positive and negative slopes implies that the test item is compared exhaustively against all list items.
A search, or sequence of comparisons, that terminated upon finding a match would finish, on average, about halfway through the list when the item was positive, but would go through the list when the test item was negative. The slope of the negative tests would then be twice that of the positives; the 2:1 slope ratio is a property of some searches of items in visible displays but not of items in memory. Thus, parallel and linear list length functions are often thought to reflect exhaustive serial search in the retrieval of information from STM.
However, linear increases in mean RT with list length are also consistent with a parallel retrieval mechanism in which all comparisons take place at the same time, but with an efficiency that depends on the number of concurrent comparisons (see Figure 2b). This is called mechanism mimicry. Mimicry of serial mechanisms by parallel mechanisms occurs when only simple measures such as the average RT are available (see Townsend and Ashby, 1983, for a mathematical treatment of this mimicry). Hence, the regularities noted by Sternberg (1975; see Figure 1a) may result from exhaustive and serial comparisons, or from a set of parallel comparisons whose efficiency is affected by the number of comparisons, or by a direct-access process (see Figure 2c) whose efficiency is affected by other factors that covary with list length. More detailed analyses discriminate the serial and parallel mechanisms. In either case, time to retrieve information from STM depends on the number of items currently being remembered.
Direct-Access in Retrieval of Item Information
A serial exhaustive search mechanism has these properties: (1) linear increases in RT with list length where (2) slopes for positive and negative tests are equal; (3) RT should not depend on the position of a positive test item on the list; (4) the fastest (minimum) RTs should increase with the length of the list; and (5) the RT variance (a measure of variation from test to test) should increase linearly with list length for both positive and negative recognition tests. Over many variants of STM item recognition experiments, properties (1) and (2) hold approximately, although RT increases with list length may be more logarithmic than linear (Briggs, 1974). Properties (3) and (4), which require more detailed breakdowns of data, fail systematically. Retrieval of item information from STM, when considered in more detail, is inconsistent with serial exhaustive processing.
RT, the time to recognize an item from STM, depends strongly on its recency (property 3 fails). Test items experienced very recently yield fast RTs, with RT increases for each less recent item. There is also a small advantage for the first list item. Recency in the study list is the controlling factor whenever rehearsal is minimal or constrained to match study order. This is easily seen when the data are graphed appropriately (i.e., Monsell, 1978; see Figure 3a). Longer lists yield slower mean RTs because they include items of less recency. Averaging over list positions yields approximately linear (or logarithmic) increases in mean RT with list length. Failure of the prediction that the minimum RT should depend fairly strongly on list length (property 4) is implied by the recency data (see Figure 3a): RT for the most recent items depends only weakly on the length of the list. Distributions of RTs from shortest to longest show only tiny shifts of the minimum with list length; average RT increases with list length largely reflect shifts in the longer RTs (see Figure 3b) (Hockley, 1984). Predictions of linear increases in variability (property 5) also fail. Other findings that contradict the exhaustive serial search mechanism are decreases in RT when a stimulus is repeated and decreased RT for stimuli with high test probability, in situations where list items are the same over long sets of trials.
The list position effects and aspects of the RT distributions in item recognition tasks contradict properties 3, 4, and 5 of the serial exhaustive scan. The approximate equality of positive and negative slopes (property 2) contradicts terminating (nonexhaustive) scan. For item memory (as distinct from order memory see subsequent explanation), the data are consistent with a direct-access retrieval mechanism in which decreased availability of less recent items determines average RT for particular list positions and hence for different list lengths (McElree and Dosher, 1989).
Time Course of STM Retrieval for Item Information
A direct-access retrieval mechanism was directly confirmed by more detailed RT methods, which allow inferences about the full time course of retrieval. These methods interrupt retrieval at various times after onset of the test display and observe the rate of increase in correct responding with additional retrieval time. B. McElree and B. Dosher (1989) showed that the rate of retrieval of items from lists of different length was fastest for the single most recent item—a case of an immediate match between the last item studied and the test item—but is otherwise unaffected by either list length or list position (see Figure 4). Retrieval from STM was parallel or direct access, yet the ultimate success of retrieval was limited by familiarity in memory. Recent items have been least affected by forgetting due to the passage of time or intervening items between study and test. The strength of items when measured by errors and the accessibility of items when measured by RT are both directly related to the recency of study, with a small additional advantage for the primacy or first item on the list.
Recognition of items presented in longer lists that exceed estimates of STM capacity shows many of the same properties as recognition of items from short, recent lists. Recognition from longer lists leads to more errors than for STM lists, where the error rates may be less than 5 percent. However, as in STM, list position is an important factor in LTM, affecting both RT and accuracy. As in shorter lists, items near the end of the list are recognized more quickly and accurately. When longer lists are used, study is usually followed by many test trials, and location in the test protocol also has a powerful effect. Earlier tests yield faster RT and accuracy. Later in the test sequence, additional time and materials are interpolated between encoding and retrieval; this is another manipulation of recency. As with the STM data, study and test position effects on average RT primarily reflect shifts in the long tail of the RT distributions. Full time course of recognition is fastest for the single most recent item, and otherwise equivalent but limited by familiarity (Wickelgren, Corbett, and Dosher, 1980). These findings rule out recency-dependent serial comparisons that terminate on a match. The details of these data, when examined carefully, are accounted for by a parallel, direct-access retrieval process with shifts in estimated familiarity (Ratcliff, 1978).
Indeed, direct access appears to be a general property of item retrieval. The same time-course patterns are found in both supra-and sub-span lists (Wickelgren, Corbett, and Dosher, 1983; McElree, 2001), indicating that direct access is a property of both short-and long-term representations. Direct-access retrieval is also evident in the recovery of information from more complex representations and based on altered cues. For example, direct-access retrieval characterizes the recognition of an item that is part of a hierarchically coded group (McElree, 1998), as well as recognition that is based on component properties (e.g., phonological and semantic properties) of the memory representation. The latter finding in particular suggests that direct access arises from a content-addressable retrieval operation that may operate either on an exact representation or via reintegration of the studied item from related cues during retrieval.
Search Processes in Recovery of Order Information
Although item information appears to be recovered through a direct-access content-addressable retrieval process, this is not so for certain other forms of information. Studies indicate that the retrieval of relational information, including temporal order information in STM and positional information in LTM often require a slow serial search. For example, if information about the order of items in STM must be retrieved, both accuracy and retrieval speed depend strongly upon item recency (McElree and Dosher, 1993). Temporal order information in STM was examined with a judgment of recency task, in which subjects were presented two test probes from a short list of five or six items and asked to select the item that occurred more recently. The accuracy and speed of the order judgment, as measured either by mean RT or by time course analysis, was directly related to the recency of the most recent item. A related phenomenon occurs in n-back tasks, in which a positive response is restricted to a particular ordinal position (e.g., a 1-back match, a 2-back match) in a long ongoing sequence instead of an overt judgment of order (McElree, 2001). In both kinds of relational or order tasks, retrieval speed is directly controlled by the recency of the most recent relevant item, strongly implicating a memory search process that begins with the most recent item. One possibility is that ordered representations in memory are serially scanned from the most recent, moving backwards in time (for specific models see McElree and Dosher, 1993; McElree, 2001). Alternatively, order information may be reconstructed at retrieval by a serial chain process (Murdock, 1982), in which the last item on the list is used as a cue to recover the next item on the list from memory, and so on. In all cases in which the retrieval of relational or order information have been evaluated, some form of (terminating) successive or serial process operates. This stands in contrast to the direct-access recovery of item information from either STM or LTM.
Relation of STM Item and Order Retrieval to STM Recall
Another classic measure of STM is the ordered recall of all items on a short list. The list length at which ordered recall is perfect 50 percent of the time (usually an interpolated value) is called the memory span. Span and the RT and accuracy for recognizing a single list item both measure aspects of STM function. Historically, span has been taken as a measure of capacity (Miller, 1956; Baddeley, 1986), while recognition RT was taken as a measure of retrieval efficiency. Although the two measures do not strongly correlate with each other across individuals, they vary together across different to-be-remembered materials. J. P. Cavanaugh (1972) compared, via a survey of the research, the memory spans and the RT list-length slopes of digits, letters, words, shapes, and nonsense materials. Materials yielding higher spans (longer list lengths supporting 50% recall) exhibit relatively shallower slopes (less increase in average RT with increasing list length; see Figure 5). The primary factor producing differences in materials in both measures of STM may be overall familiarity (Puckett and Kausler, 1984).
The relationship between item recognition and memory span in ordered recall may reflect the different retrieval demands of the two tasks. Item recognition reflects the strength of direct-access content addressable information in memory, while span, a measure of ordered recall, additionally requires the sequential access of ordered representations, either by sequential access of an intrinsically ordered representation, or by chained retrieval from a content-addressable store. In support of this view that STM recall is composed of a sequence of retrieval or scanning operations, the time to complete ordered recall of a span length list may be as long as 5 to 8 s (Dosher and Ma, 1998), generally compatible both with the time course of individual retrievals (see Figure 4) and with the time course of recovery of item information.
Relation to Other Abilities
The capacity of STM has been viewed as an elementary information-handling process, related to efficiency in a variety of mental tasks (Baddeley, 1986). The speed of access to information in STM, defined as the slope of the dependence of RT on list length (s in the linear equation above), has been tested as a correlate of the quality of performance on general cognitive indices such as aptitude scores. Correlations of capacity with psychometric measures are usually higher than those of retrieval time with psychometric measures (Sternberg, 1975). However, various special populations, such as the young and the elderly, have been shown to have characteristic increases in STM recognition times, either in base times or in slopes, compared with the performance of young adults (Sternberg, 1975). STM function is one of the important information-processing correlates with verbal intelligence (Palmer, MacLeod, Hunt, and Davidson, 1985). The concepts of STM and its close relative, working memory, have been central in recent years in the development of theories and predictive indices of general cognitive functions (Engle, Tuholski, and Kane, 1999).
Briggs, G. E. (1974). On the predictor variable for choice reaction time. Memory & Cognition 2, 575-580.
Burrows, D., and Okada, R. (1975). Memory retrieval from long and short lists. Science 188, 1,031-1,033.
Cavanaugh, J. P. (1972). Relation between the immediate memory span and the memory search rate. Psychological Review 79, 525-530.
Dosher, B., and Ma, J. J. (1998). Output loss or rehearsal loop? Output time versus pronunciation time limits in immediate recall for forgetting matched material. Journal of Experimental Psychology: Learning, Memory, and Cognition 24, 316-335.
Engle, R. W., Tuholski, S., and Kane, M. (1999). Individual differences in working memory capacity and what they tell us about controlled attention, general fluid intelligence and functions of the prefrontal cortex. In A. Miyake and P. Shah, eds., Models of working memory: Mechanisms of active maintenance and executive control. Cambridge, MA: Cambridge University Press.
Gillund, G., and Shiffrin, R. M. (1984). A retrieval model for both recognition and recall. Psychological Review 91, 1-67.
Hockley, W. E. (1984). Analysis of response time distribution in the study of cognitive processes. Journal of Experimental Psychology: Learning, Memory, and Cognition 10, 598-615.
James, W. (1890; reprint 1950). The principles of psychology. New York: Dover.
McElree, B. (1998). Attended and non-attended states in working memory. Journal of Memory and Language 38, 225-252.
—— (2001). Working memory and focal attention. Journal of Experimental Psychology: Learning, Memory, and Cognition 27, 817-835.
McElree, B., and Dosher, B. (1989). Serial position and set size in short-term memory: The time course of recognition. Journal of Experimental Psychology: General 118, 346-373.
—— (1993). Serial retrieval processing in the recovery of order information. Journal of Experimental Psychology: General 122, 291-315.
Miller, G. E. (1956). The magic number seven, plus or minus two: Some limits of our capacity for processing information. Psychological Review 63, 81-97.
Monsell, S. (1978). Recency, immediate recognition memory, and reaction times. Cognitive Psychology 10, 465-501.
Murdock, B. B., Jr. (1971). A parallel processing model of scanning. Perception and Psychophysics 10, 289-291.
—— (1982). A theory for the storage and retrieval of item and associative information. Psychological Review 89, 609-626.
Palmer, J., MacLeod, C. M., Hunt, E., and Davidson, J. E. (1985). Information processing correlates of reading. Journal of Memory and Language 24, 59-88.
Puckett, J. M, and Kausler, D. H. (1984). Individual differences and models of memory span: A role for memory search rate? Journal of Experimental Psychology: Learning, Memory, and Cognition 10, 72-82.
Raaijmakers, J. G., and Shiffrin, R. M. (1981). Search of associative memory. Psychological Review 88, 93-134.
Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review 85, 59-108.
Sternberg, S. (1966). High speed scanning in human memory. Science 153, 652-654.
—— (1975). Memory-scanning: New findings and current controversies. Quarterly Journal of Experimental Psychology 27, 1-32.
Townsend, J. T., and Ashby, F. G. (1983). The stochastic modeling of elementary psychological processes. New York: Cambridge University Press.
Wickelgren, W. A., Corbett, A. T, and Dosher, B. A. (1980). Priming and retrieval from short-term memory: A speed-accuracy tradeoff analysis. Journal of Verbal Learning and Verbal Behavior 19, 387-404.