Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org by California State University - Channel Islands on 09/12/12. For personal use only. Working Memory: Theories, Models, and Controversies

Alan Baddeley

Department of Psychology, University of York, York YO10 5DD, United Kingdom; email: [email protected]

Annu. Rev. Psychol. 2012. 63:1–29 Keywords First published online as a Review in Advance on attention, central executive, phonological loop, episodic buffer, Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org September 27, 2011 by California State University - Channel Islands on 09/12/12. For personal use only. visuo-spatial sketchpad, short-term memory The Annual Review of Psychology is online at psych.annualreviews.org Abstract This article’s doi: I present an account of the origins and development of the multicom- 10.1146/annurev-psych-120710-100422 ponent approach to working memory, making a distinction between Copyright c 2012 by Annual Reviews. the overall theoretical framework, which has remained relatively stable, All rights reserved and the attempts to build more specific models within this framework. 0066-4308/12/0110-0001$20.00 I follow this with a brief discussion of alternative models and their rela- tionship to the framework. I conclude with speculations on further de- velopments and a comment on the value of attempting to apply models and theories beyond the laboratory studies on which they are typically based.

1 Contents THE EPISODIC BUFFER ...... 15 WORKING MEMORY: WMandBinding...... 15 THEORIES, MODELS, AND Visual Binding and WM ...... 16 CONTROVERSIES ...... 2 Binding in Verbal WM ...... 17 Short-Term Memory ...... 4 LINKING LONG-TERM AND Evolution of a Multicomponent WORKING MEMORY...... 18 Theory...... 5 Is WM Just Activated LTM? . . . 18 CHARACTERISTICS OF THE Long-TermWM...... 18 PHONOLOGICAL LOOP . . . 7 LTM and the Multicomponent The Phonological Similarity Model...... 18 Effect...... 8 NEUROBIOLOGICAL TheWordLengthEffect...... 8 APPROACHES TO Articulatory Suppression ...... 8 WORKING MEMORY...... 18 Irrelevant Sound Effects ...... 8 SOME ALTERNATIVE Retaining Serial Order...... 9 APPROACHES...... 19 The Phonological Loop TheoriesofSTM...... 19 andLTM...... 10 TheoriesofWM...... 19 The Phonological Loop: Cowan’s Embedded Processes Master or Slave?...... 11 Theory...... 20 The Phonological Loop: Individual Difference–Based Critique ...... 11 Theories...... 20 VISUO-SPATIAL Jonides and the Mind SKETCHPAD...... 12 andBrainofSTM...... 21 VisualSTM...... 12 Computational Models SpatialSTM...... 12 ofWM...... 21 Visuo-SpatialWM...... 13 WHATNEXT?...... 22 THE CENTRAL A Speculative Model EXECUTIVE...... 13 and Some Questions...... 22 The Executive as Integration...... 23 Homunculus ...... 13 In Praise of Negative Results. . . 24 Fractionating the Executive . . . . 14 Applications...... 24 Interfacing with LTM ...... 14 CONCLUSION...... 25 Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org by California State University - Channel Islands on 09/12/12. For personal use only.

WORKING MEMORY: THEORIES, since the early years, not all of course related MODELS, AND CONTROVERSIES to or supportive of my own work, but a recent attempt to review it (Baddeley 2007) ended I was honored, pleased, and challenged by with more than 50 pages of references. What the invitation to write this prefatory chapter, follows is a partial, as opposed to impartial, pleased because it offered the chance to take a WM: working account of the origins of the concept of multi- broad and somewhat autobiographical view of memory component working memory (M-WM) and of my principal area of interest, working memory M-WM: my own views on its subsequent development. (WM), but challenged by the potential magni- multicomponent My first draft would have filled the chapter tude of the task. The topic of working memory working memory page allowance with references; I apologize to has increased dramatically in citation counts

2 Baddeley all of those whose work should have been cited rival theories to confront each other in the and is not. all-important “crucial experiment” that settles I entered psychology as a student at Univer- the issue. This approach was closer in spirit to sity College London in 1953, a very exciting Hull than to Tolman. time for the field of psychology, which had My own first published study (Baddeley benefited greatly from developments during 1960) attempted just such a crucial experiment, the Second World War, where theory was en- predicting that rats would be smarter than they riched by the need to tackle practical problems. should be according to Hullian theory, and As a result, prewar issues such as the conflict demonstrating, to my own satisfaction at least, between Gestalt psychology and neobehavior- that this was the case. Alas, by the time it was ism began to be challenged by new data and published, the whole field of learning theory new ideas, some based on cybernetics, the study seemed to have collapsed. Neither side was able of control systems, with others influenced by to deliver a knockout blow, and people simply the newly developed digital computers. This in abandoned the research area. I resolved at that turn led to a renewed interest in the philosophy point that if I myself were to develop a theory, of science as applied to psychology. Typical it would be based very closely on the evidence, questions included, is psychology a science?; if which would survive even if the theory proved so, is it cumulative or are we doomed to keep totally wrong. It is an approach I have followed on asking the same questions, as appeared to ever since. be the case in philosophy? What would a good But what is the answer to our original ques- psychological theory look like? tion, should theorists be architects, building As students we were offered two answers to elegant structures such as Newton did, or this question. The first, championed by Cam- should they be explorers, gradually extending bridge philosopher Richard Braithwaite (1953), the theory on the basis of more and more regarded Newton’s Principia as the model to evidence, as in the case of Darwin? Clearly both which scientific theories should aspire, involv- Newton and Darwin got it right, but for fields ing as it does postulates, laws, equations, and at a different stage of development. Newton predictions. Within psychology, the Newto- claimed that his success resulted from “standing nian model was explicitly copied by Clark Hull on the shoulders of giants,” who no doubt stood in his attempt to produce a general theory of on the shoulders of lesser mortals like ourselves. learning, principally based on the study of maze Darwin had few such giants available. I suggest learning in the albino rat. that any complete theory is likely to require An alternative model of theorizing came explorers in its initial stages and architects to from Oxford, where Stephen Toulmin (1953) turn the broad concepts into detailed models. I argued that theories were like maps, ways myself am very much at the explorer end of the Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. of organizing our existing knowledge of the continuum, but I fully accept the importance of world, providing tools both for interacting with the skills of the architect if theory is to develop. the world and for further exploration. Edward My research career really began with my ar- Tolman in Stanford had a view of learning in rival at the Medical Research Council Applied rats that fitted this model, using it to challenge Psychology Unit (APU) in Cambridge. Its role Hull’s neo-behaviorist approach. This raised was to form a bridge between psychological the- the crucial question as to how you might decide ory and practical problems, and the year I ar- between the two apparently opposing views. rived, Donald Broadbent, its director, had just The dominant answer to that question, in the published his seminal book, Perception and Com- United Kingdom at least, was provided by Karl munication, which provided one of the sparks Popper (1959), a Viennese-trained philosopher that ignited what subsequently became known who argued strongly that a valid theory should as the cognitive revolution. I was assigned to make clear, testable predictions, allowing the work on optimizing the design of postal codes,

www.annualreviews.org • Working Memory: Theories, Models, and Controversies 3 which led me to combine the classic tradition of I decided to see if the acoustic similarity ef- nonsense syllable learning with new ideas from fect could be used to provide sensitive indirect information theory, resulting in my generat- measure of telephone line quality. It did not; STM: short-term memory ing memorable postal codes for each town in the effects of noise and similarity were simply the United Kingdom. The Post Office thanked additive, but I was intrigued by the sheer mag- LTM: long-term memory me and went on their way regardless; the code nitude of the similarity effect. Similarity was a they adopted could, however, have been much central variable within the dominant stimulus- worse, as is indeed the case in some countries, response interference theory of verbal learning but that is another story. (see Osgood 1949), but the type of similarity By this time my approach to theory was seemed not to be regarded as important. So, evolving away from Popper’s idea of the need would Conrad’s effect generalize to other types for crucial experiments, largely on the grounds of similarity in STM? that clear predictions only appeared to be pos- I tested this, comparing recall of sequences sible in situations that were far narrower than with five phonologically similar words (man, the ones I found interesting. I subsequently dis- mat, can, map, cat), five dissimilar words (e.g., covered that within the philosophy of science, pit, day, cow, pen, sup), and five semantically sim- Lakatos (1976), and allegedly Popper himself, ilar sequences (huge, big, wide, large, tall ) with had subsequently abandoned the reliance on five dissimilar (wet, soft, old, late, good ). I found falsification, arguing instead that the mark of a (Baddeley 1966a) a huge effect of phonological good theory is that it should be productive, not similarity1 (80% sequences correct for dissimi- only giving an account of existing knowledge, lar, 10% for similar) and a small but significant but also generating fruitful questions that will effect for semantic similarity (71% versus increase our knowledge. This more map-like 65%). I went on to demonstrate that this pat- view of theory is the one that I continue to take. tern reversed when long-term memory (LTM) was required by using ten-word lists and several Short-Term Memory learning trials; semantic similarity then proved The term “working memory” evolved from the critical (Baddeley 1966b). I concluded that earlier concept of short-term memory (STM), there were two storage systems, a short-term and the two are still on occasion used inter- phonological and a long-term semantically changeably. I will use STM to refer to the sim- based system. My telephony project was passed ple temporary storage of information, in con- on to a newly arrived colleague and I was left trast to WM, which implies a combination of free to explore this line of basic research. storage and manipulation. I saw my work as fitting into a pattern of My interest in STM began during my time evidence for separate STM and LTM stores. at the APU in Cambridge and was prompted Other evidence came from amnesic patients Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. by an applied problem, that of finding a way of who had preserved STM and impaired LTM, evaluating the quality of telephone lines that while other patients showed the reverse pattern might be more effective than a simple listening (Shallice & Warrington 1970). A third source of test. My PhD supervisor Conrad had recently evidence came from two-component memory discovered the acoustic similarity effect. He tasks, which comprised a durable LTM com- was studying memory for proposed telephone ponent together with a temporary component. dialing codes when he noted that even with A typical example of this was the recency effect visual presentation, memory errors resembled v acoustic mis-hearing errors (e.g., for b), and 1I subsequently abandoned the term “acoustic similarity” be- that memory for similar sequences (bgtp cause it suggested an input modality-based system, which is c) was poorer than for dissimilar (krlqy), not the case; I mistakenly assumed that phonological was a more neutral term. It was not intended as a statement of the concluding that STM depends on an acoustic linguistic basis of the memory system, which remains an open code (Conrad & Hull 1964). question.

4 Baddeley in free recall (Glanzer 1972); the last few words & Norman (1965) and by Atkinson & Shiffrin of a list are well recalled on immediate test but (1968), but it has often been neglected in subse- not after a brief filled delay, unlike earlier items. quent years. Material tested after a brief delay At this point, my simple assumption of two (i.e., an STM task) is likely to reflect both LTM stores, with STM phonologically based and and some form of temporary storage. LTM semantically based, led to some clear pre- dictions. Amnesic patients should have seman- Evolution of a Multicomponent tic coding problems, and recency should be Theory acoustically based. Studies based on amnesic pa- After nine years at the APU, I moved to tients suffering from Korsakoff’s syndrome did Sussex into a new department of experimental suggest a semantic encoding deficit (Cermak psychology, where, in 1972, I was joined by et al. 1974), but our own work showed no evi- Graham Hitch as a post-doctoral fellow on dence of such a deficit (Baddeley & Warrington my first research grant. After a first degree 1970), and later work (Cermak & Reale 1978) in physics, he had done a psychology MSc attributed their previously observed deficit to in Sussex and a PhD with Broadbent at the additional executive problems, often found in APU. We had proposed (perhaps unwisely) to Korsakoff’s syndrome. investigate the link between STM and LTM, In the case of two-component tasks, it be- beginning our grant just when the previously came clear that recency did not depend on ver- popular field of STM was downsizing itself bal STM (Baddeley & Hitch 1977) and that following criticism of the dominant Atkinson the use of semantic or phonological coding & Shiffrin (1968) model for three reasons. was strategy dependent. Phonological coding First, the model assumed that merely holding of verbal material is rapid, attentionally un- information in STM would guarantee transfer demanding, and very effective for storing se- to LTM, whereas Craik & Lockhart (1972) rial order. Semantic coding can be rapid for showed that the nature of processing is crucial, meaningful sequences such as sentences, but with deeper, more elaborate processing leading it is much harder to use for storing the order to better learning. Second, its assumption that of unrelated words (Baddeley & Levy 1971). the short-term store was essential for access to We also showed that word sequences can si- LTM proved to be inconsistent with neuropsy- multaneously be encoded both phonologically chological evidence. Patients with a digit span and semantically (Baddeley & Ecob 1970) and of only two items and an absence of recency in that standard tasks such as immediate serial re- free recall should, according to Atkinson and call can reflect both long-term and short-term Shiffrin, have a defective short-term store that components, each of which may be influenced should lead to impaired LTM. This was not the by either phonological or semantic factors. In case. Third, given that Atkinson and Shiffrin Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. short, STM, retention of material over a brief assumed their short-term store to be a working period, may be based on either phonological or memory, playing an important general role semantic coding. The former is easy to set up in cognition, such patients should have major but readily forgotten; the latter may take longer intellectual deficits. They did not. One patient, to set up but tends to be more durable. Both can for instance, was an efficient secretary, and operate over brief delays, and the fact that we another ran a shop and a family. Interest in the can learn new words indicates that long-term field began to move from STM to LTM, to phonological learning also occurs. semantic memory and levels of processing. It is worth emphasizing the need to distin- Graham Hitch and I did not have access guish between STM as a label for a paradigm in to these rare but theoretically important which small amounts of information are stored STM-deficit patients and instead decided over brief delays and STM as a theoretical stor- that we would try to manufacture our own age system. This point was made by Waugh “patients” using student volunteers. We did

www.annualreviews.org • Working Memory: Theories, Models, and Controversies 5 so, not by removing the relevant part of their systems, all of which were limited in capacity. brain, but by functionally disabling it by We labeled the central controller as a “central requiring participants to do a concurrent task executive” (CE), initially referring to the ver- CE: central executive that was likely to occupy the limited-capacity bal system as the “articulatory loop,” after the short-term storage system to varying degrees. subvocal rehearsal assumed to be necessary to The concurrent task we chose was serial maintain information, and later adopting the verbal recall of sequences of spoken digits. As term “phonological loop” to emphasize stor- sequence length increased, the digits should age rather than rehearsal. We termed the third occupy more and more of available capacity, component the “visuo-spatial sketchpad,” leav- with the result that performance on any ing open the issue of whether it was basically task relying on WM should be progressively visual, spatial, or both. impaired. In one study, participants performed We began by focusing on the phonological a visually presented grammatical reasoning loop on the grounds that it seemed the most task while hearing and attempting to recall tractable system to investigate, given the very digit sequences of varying length. Response extensive earlier research on verbal STM. At time increased linearly with concurrent digit this point, I unexpectedly received an invitation load. However, the disruption was far from from Gordon Bower to contribute a chapter to catastrophic: around 50% for the heaviest load, an influential annual publication presenting re- and perhaps more strikingly, the error rate cent advances in the area of learning and mem- remained constant at around 5%. Our results ory. We hesitated; our model was far from com- therefore suggested a clear involvement of plete, should we perhaps wait? We went ahead whatever system underpins digit span, but not anyhow (Baddeley & Hitch 1974), presenting a a crucial one. Performance slows systematically model that is still not complete nearly 40 years but does not break down. We found broadly and many publications later. similar results in studies investigating both Over the next decade we continued to verbal LTM and language comprehension, and explore the model and its potential for ap- on the basis of these, abandoned the assump- plication beyond the cognitive laboratory. At tion that WM comprised a single unitary store, this point I agreed to summarize our progress proposing instead the three-component system in a monograph (Baddeley 1986). This was shown in Figure 1 (Baddeley & Hitch 1974). approaching completion when I realized that We aimed to keep our proposed system as I had said nothing about the CE, very much simple as possible, but at the same time, po- a case of Hamlet without the prince. My re- tentially capable of being applied across a wide luctance to tackle the executive stemmed from range of cognitive activities. We decided to two sources: first, its probable complexity, and split attentional control from temporary stor- second, because of the crucial importance of Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. age, which earlier research suggested might rely its attentional capacity. Although there were a on separate verbal and visuo-spatial short-term number of highly developed and sophisticated theories of attention, most were concerned with the role of attention in perception, whereas the principal role of the CE was the attentional control of action. The one directly relevant article I could find (Norman & Shallice 1986) appeared as a chapter because of the difficulty of persuading a journal to accept it (Shallice 2010, personal communication), alas, all too common with papers presenting new ideas. Figure 1 Norman and Shallice proposed that action The original Baddeley & Hitch (1974) working memory model. is controlled in two rather separate ways. One is

6 Baddeley based on well-learned habits or schemata, de- is the assumption that each of these systems can manding little in the way of attentional con- be fractionated into subsystems and that these trol. An example of this might be the activity will be linked to perceptual and LTM processes of driving a well-learned route to your office. in ways that require further investigation. This source of control can be overridden by a My overall view of WM therefore com- second process, the supervisory attentional sys- prised, and still comprises, a relatively loose tem (SAS), which responds to situations that theoretical framework rather than a precise are not capable of being handled by habit-based model that allows specific predictions. The processes, for example, coping with the closure success of such a framework should be based, of a road on your normal route. as suggested by Lakatos (1976), not only on With some relief, I incorporated the its capacity to explain existing data but also on Norman and Shallice model into my own con- its productivity in generating good, tractable cept of a CE, producing a book (Baddeley 1986) questions linked to empirical methods that can that attempted to pull together developments in be widely applied. The proposed components WM that had occurred in the previous decade of WM are discussed in turn, beginning with and then apply them to data from the literature the phonological loop. in three areas: fluent reading, the development of WM in children, and the effects of aging. Although I tended to refer to our proposals as CHARACTERISTICS OF THE a model, using the criteria proposed earlier, it PHONOLOGICAL LOOP might better be regarded as a simple theory, We saw the phonological loop as a relatively in the sense of Toulmin’s idea of theories as modular system comprising a brief store to- maps, linking together existing knowledge and gether with a means of maintaining information encouraging further investigation. If so, it was a by vocal or subvocal rehearsal. In the 1960s, a map with many blank areas that I hoped would number of studies attempted to decide whether be filled by myself and others, leading in due forgetting in the STM system was based on course to more detailed modeling. trace decay or interference (see Baddeley 1976). What then are the essentials of the broad None of these studies proved to be conclusive, theory? The basis is the assumption that it a state of affairs that remains true, in my own is useful to postulate a hypothetical limited- opinion. We opted to assume a process of trace capacity system that provides the temporary decay, partly on the basis of our results and storage and manipulation of information that is partly because it avoided the need to become in- necessary for performing a wide range of cogni- volved in the many controversies surrounding tive activities. A second assumption is that this traditional approaches to interference theory system is not unitary but can be split into an ex- at the time (see Baddeley 1976, chapter 5), al- Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. ecutive component and at least two temporary though we did assume a limited-capacity store, storage systems, one concerning speech and which in turn implies some unspecified form sound while the other is visuo-spatial. These of interference, either by displacement or by three components could be regarded as mod- overwriting. We used existing results, together ules in the sense that they comprise processes with our own subsequent studies, to create a and storage systems that are tightly interlinked simple model that is based on the method of within the module and more loosely linked converging operations. This involves combin- across modules, with somewhat more remote ing evidence from a range of different phenom- connections to other systems such as percep- ena, each consistent with the model, but each tion and LTM. I regard the very rigid defini- individually explicable in other ways. If none tion of modularity by Fodor (1983) as unhelpful of the competing interpretations are able to ex- and neuropsychologically implausible. A conse- plain the whole pattern, whereas the phonolog- quence of my rejection of Fodorian simplicity ical loop model can, then this provides valuable

www.annualreviews.org • Working Memory: Theories, Models, and Controversies 7 support. This approach has the advantage of since faced challenge and counter-challenge potentially producing a robust model, but it has (see Baddeley 2007, pp. 43–49). Fortunately, the disadvantage of being required to confront however, the general hypothesis of a phono- a range of different possible alternative expla- logical loop will function equally well with nations for each individual phenomenon. either a decay or interference interpretation of short-term forgetting, illustrating the value The Phonological Similarity Effect of combining a broad theoretical map while leaving more detailed modeling to be decided As described above, this is regarded as an in- by further experimentation. dication that phonological storage is involved. Its effect is principally on the storage of order information. Indeed, item information may be Articulatory Suppression helped by similarity since it places constraints If the word length effect is dependent on on possible responses. For this reason, studies subvocalization, then preventing it should that specifically attempt to investigate the loop eliminate the effect. This is indeed the case tend to minimize the need to retain item in- (Baddeley et al. 1975b). When participants formation by repeatedly using the same lim- are required to continuously utter a single ited set, for example, consonants. Studies using word such as “the,” performance drops and is open sets, for instance, different words for each equivalent for long and short words. Suppres- sequence, are more likely to reflect loss of item sion also removes the phonological similarity information and to show semantic and other effect for visually presented materials but not LTM-based effects. when presentation is auditory (Baddeley et al. 1984). We interpret this as suggesting that The Word Length Effect spoken material gains obligatory access to the phonological store, whereas written material We assumed that vocal or subvocal rehearsal needs to be subvocalized if it is to register. was likely to occur in real time, with longer The claim that auditory presentation allows words taking longer and hence allowing more a phonological trace to be laid down despite time for trace decay, thus leading to poorer suppression has recently been challenged. Jones performance. We studied the immediate recall et al. (2006) have suggested that the effect is lim- of sequences of five words ranging in length ited to the recency component of immediate se- from one syllable (e.g., pen day hot cow tub)to rial recall, suggesting that it is better regarded five syllables (e.g., university, tuberculosis, oppor- as a perceptual effect. However, although this tunity, hippopotamus, refrigerator) and found that may be true for long lists, shorter lists show an performance declined systematically with word effect that operates throughout the serial posi- length. As expected, when participants were re- Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org tion curve (Baddeley & Larsen 2007). by California State University - Channel Islands on 09/12/12. For personal use only. quired to read out words of different lengths as rapidly as possible, there was a close corre- spondence between word length and articula- Irrelevant Sound Effects tion time. The simple way of expressing our Colle & Welsh (1976) required their partici- results was to note that people are able to re- pants to recall sequences of visually presented member as many words as they can articulate in digits presented either in silence or accompa- two seconds (Baddeley et al. 1975b). nied by white noise or by speech in an unfamil- We interpreted our data by assuming that iar language that they were told to ignore. Only longer words take longer to rehearse, resulting the spoken material disrupted performance on in more trace decay and poorer recall. Such de- the visually presented digits, an effect that was cay is also likely to continue during the slower independent of the loudness of the irrelevant spoken recall of longer words. We presented sound sources. Pierre Salame, a French visitor evidence for time-based decay, which has to Cambridge, and I followed up and extended

8 Baddeley Colle’s work, demonstrating that visual STM forgotten, but rather because their order is was disrupted to the same extent by irrelevant lost. Retaining serial order is a crucial demand words and nonsense syllables; indeed, irrele- for a wide range of activities, notably includ- vant digits had no more effect on digit recall ing language, in which sequences of sounds than did nondigit words containing the same within words and words within sentences must phonemes (e.g., one two replaced by tun woo), be maintained, and skilled motor performance suggesting that interference was operating at a such as striking a ball with a bat or playing the prelexical level. We did, however, find slightly piano. However, as Lashley (1951) points out, it less disruption of our monosyllabic digits from is far from easy to explain how this is achieved. bisyllabic words than from monosyllabic words, The most obvious hypothesis is through the concluding rather too hastily that this suggested previously described mechanism of chaining that interference was dependent on phonolog- through sequential associations. However, this ical similarity (Salame & Baddeley 1986). Like has some major potential problems; if one item Colle and Welsh, we suggested an interpreta- is lost, then the chain is broken and subsequent tion in terms of some form of mnemonic mask- recall should fail, and yet it is often the case that ing. This proved to be something of an embar- despite errors in the middle of a sequence, the rassment when it was clearly demonstrated that latter part is reproduced correctly. Similarly, if irrelevant items that were phonemically similar an item is repeated within the chain (e.g., 753 to the remembered sequence were no more dis- 596), then the chain should be disrupted, but ruptive than dissimilar items ( Jones & Macken this disruption, when it occurs, is typically far 1995, Larsen et al. 2000). Unfortunately, our from dramatic. initial hypothesis came to be regarded as cen- A third phenomenon appears to be even tral to WM, despite our subsequent withdrawal, more problematic. This again is an effect that a salutary lesson in premature theorizing. was discovered when trying to solve a practi- Meanwhile Dylan Jones and colleagues in cal problem, that of trying to reduce the neg- Wales were developing a very extended pro- ative impact of phonological similarity on the gram of research on irrelevant sound. They recall of postal codes. It seemed plausible to me showed that STM was disrupted not only to assume that the principal effect of similar- by irrelevant speech, but also by a range of ity would come from having two or more sim- other sounds, including, for example, fluctu- ilar items bunched together, in which case it ating tones ( Jones & Macken 1993). In order might prove possible to greatly minimize the ef- to account for their results they proposed the fect by alternating similar and dissimilar items “changing state” hypothesis, whereby the cru- (e.g., dfvkpl ). The results were disappointing; cial feature was that the irrelevant sound needed the similar items appeared to be just as liable to fluctuate. Jones (1993) coupled this with the to be forgotten when sandwiched between dis- Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. object-orientated episodic record (OOE-R) hy- similar items as when they were adjacent, so pothesis, which assumes that both digits and ir- we put the experiment to one side. It was only relevant sounds are represented as potentially later, when I was attempting to pin down the competing paths on a multidimensional surface. nature of the phonological loop effect, that I The OOE-R hypothesis is not spelled out in de- realized that our result had clear implications tail but would appear to assume that serial order for theories of serial order retrieval in general is based on chaining, whereby each item acts as (Baddeley 1968) and were in particular incon- a stimulus for the response that follows, which sistent with hypotheses that depended upon in turn acts as a further stimulus. chaining. The argument goes as follows: If one considers a sequence of six letters as a series of Retaining Serial Order pairs, then we know that the principal source of A typical memory span is around six or seven interference comes from similarity at the stim- digits, not because the digits themselves are ulus level, which then gives rise to errors on the

www.annualreviews.org • Working Memory: Theories, Models, and Controversies 9 subsequent response (Osgood 1949). We would generalized to visual STM? The answer seems therefore expect errors to follow the similar to be that they can (Hurlstone 2010). If so, do items, whereas in fact the similar items them- they reflect a single common system? I myself selves were the main source of error (Baddeley think it more likely that evolution has applied 1968). This result has continued to present a the same solution to a problem, maintaining challenge to models of serial order. serial order, that crops up in a range of different The past decade has seen considerable activ- domains. ity in the attempt to produce clearly specified computational or mathematical models of serial order retention, with a number located within The Phonological Loop and LTM the phonological loop tradition. Very briefly, What function might the phonological loop approaches fall into two categories. One class of (PL) serve, other than making telephoning models assumes that items are associated with a easier (an unlikely target for Mother Nature)? series of internal markers, which may be tempo- The opportunity to investigate this question ral oscillators as in Brown et al.’s (2000) OSCAR cropped up when an Italian colleague, Giuseppe hypothesis, or other forms of ordinal marking, Vallar, invited me to help him to investigate as in the case of the model and its subsequent a patient, PV, with a very pure and specific refinement by Burgess & Hitch (1999, 2006). deficit in phonological STM. Her intellect was A second approach is typified by the primacy preserved, but her auditory digit span was only hypothesis of Page & Norris (1998), which as- two items. She had fluent language production sumes a limited capacity of excitation that is and comprehension, except for long, highly shared among the sequence of items. The first artificial sentences in which ambiguity could item is the most strongly activated, the second only be resolved by retaining the initial part of slightly less, and so forth. At recall, the strongest a long sentence until the end, again not a great item is retrieved first and then inhibited to avoid evolutionary gain. We then came up with the further repetition before going on to the next idea that her phonological loop might be neces- strongest. Both of these approaches can handle sary for new long-term phonological learning. the similarity sandwich effect, as they do not de- We tested this by requiring her to learn Russian pend upon chaining. Furthermore, they require vocabulary (e.g., flower-svieti ), comparing this two stages, a store and a serial order link, of- with her capacity for learning to pair unrelated fering an interpretation of the irrelevant sound Italian words, for example (castle-table). When effect in terms of adding noise to this additional compared to a group of matched controls, stage (Page & Norris 2003), an explanation as to her capacity to learn native language pairs was why similarity between irrelevant and remem- normal, whereas she failed to learn a single bered items is not important. Russian word after ten successive trials, a point Annu. Rev. Psychol. 2012.63:1-29. Downloaded from www.annualreviews.org

by California State University - Channel Islands on 09/12/12. For personal use only. Modeling serial order continues to be a at which all the normal participants had perfect very lively field with considerable interaction performance (Baddeley et al. 1988). We had between proponents of the different models, found a function for the phonological loop. which are now starting to become more ambi- Although the work with PV had a major tious. Burgess and Hitch are now attempting to influence on my theoretical views, of much model the link between the phonological loop greater practical importance was my collabo- and long-term phonological learning (Burgess ration with Susan Gathercole, in which we ex- & Hitch 2006, Hitch et al. 2009), while a plored the role of the phonological loop in vo- further challenge being addressed lies in the cabulary learning, both in children with specific interpretation of chunking, the effect that language impairment and in normal children. A makes sentences so much more readily recalled series of studies showed that WM plays a signif- than scrambled words (Baddeley et al. 2009). icant role in the initial stages of vocabulary ac- Can models of serial order in verbal STM be quisition and is also linked to reading skills (see

10 Baddeley Baddeley et al. 1998 for a review). It formed the The Phonological Loop: basis of an extensive and successful application Master or Slave? of the M-WM theory to the identification and In formulating our model, we referred to the treatment of WM deficits in school-age chil- loop and sketchpad as slave systems, borrow- dren (Gathercole & Alloway 2008; Gathercole ing the term from control engineering. It is, et al. 2004a,b). however, becoming increasingly clear that the At a theoretical level, work with PV led to a loop can also provide a means of action con- major development. I had previously tended to trol. In my own case, this first became obvious treat WM and LTM as separate though interre- during a series of studies of the CE, in this case lated systems. The fact that the loop specifically concentrating on its capacity for task switch- facilitates new phonological learning implies a ing. We used a very simple task in which par- direct link from the loop to LTM. Gathercole ticipants were given a column of single digits (1995) showed that existing language habits in- and required in one condition to add 1 and fluence immediate nonword recall, making the write down a total, and in another condition, nonw