Short-Term or Working Memory 437presumably of integrating phonological, visual, and possibly
other types of information” (Baddeley, 2000, p. 419).
To solve these and other problems, Baddeley recently pro-
posed a new working memory component—the episodic
buffer—to serve as a limited capacity temporary storage sys-
tem (see Figure 15.4). Controlled by the central executive,
the episodic buffer differs from the loop and the sketchpad in
performing both a storage and an integrative function; infor-
mation from many different sources can be tied together in
the buffer, including semantic information, and the result is a
multidimensional episodic code. Presumably, the buffer en-
ables one to store material when the loop or the sketchpad is
unavailable, and it helps to explain how certain item charac-
teristics (e.g., lexicality or imageability) might affect remem-
bering over the short term. It is difficult to judge the merits of
this new component of working memory at this point, how-
ever, because it has little, if any, unique empirical support.
Simulation Models of Short-Term Memory
Over the years, a number of formal simulation models (either
mathematical or computer-based) have been proposed to ex-
plain the particulars of short-term retention. Probably the best
known is the Atkinson and Shiffrin (1968) buffer model,
which maintained a distinction between the structural fea-
tures of a memory system (e.g., a limited-capacity short-term
store) and the strategic control processes that operate within
those structures (e.g., rehearsal, coding, or both). The buffer
model established the mold for many subsequent modeling
attempts, but most current models possess a decidedly differ-
ent flavor. For example, whereas little attention was given in
the Atkinson and Shiffrin model to the retrieval of short-term
memories (items were simply dumped out of the short-term
store), most current efforts focus extensively on the retrieval
and interpretation of activity traces.
Virtually all current models maintain the distinction be-
tween short- and long-term memory, but they differ in whether
similar processes are assumed to operate in the two cases.
For example, some models essentially mimic the standard jug-
gler model and attribute many of the standard immediate
memory phenomena to a trade-off between rehearsal and
decay (e.g., Burgess & Hitch, 1999; Henson, 1998; Page &
Norris, 1998). Other so-called unitary models reject the con-
cept of decay and offer little role for rehearsal, assuming in-
stead that short-term retention is controlled by the same
processes that control all forms of remembering—that is,
both short- and long-term (e.g., Anderson & Matessa, 1997;
G. D. A. Brown, Preece, & Hulme, 2000; Nairne, 1990). Space
does not permit a full accounting of these models, but I very
briefly outline some of their main features in the following two
sections.Hybrid ModelsI use the term hybridto classify the first set of models, be-
cause most acknowledge the important role that cue-driven
retrieval processes play in short-term retention. Thus, items
are not dumped out of a short-term buffer or loop; instead, re-
call candidates are chosen through some kind of item selec-
tion mechanism. This is a characteristic shared by unitary
models that assign no special properties to remembering over
the short term. In most other respects, though, hybrid models
are simply implementations of the standard juggler model:
Performance is based on short-term activity traces that are
subject to immediate decay in the absence of continued inter-
nal rehearsal.
In the primacy model of Page and Norris (1998), imme-
diate retention of serial order is controlled by the relative
activation levels of list item traces. Activation level is deter-
mined by a primacy gradient, such that the trace for the first
list item is assumed to be more active than the second list
item, and so on. At the point of recall, items are selected for
output based on their activation level, which means that the
first list item tends to be output first and then suppressed. The
output selection process is noisy, so there is a certain proba-
bility that items will be selected for output out of their proper
sequence (leading to errors). Page and Norris (1998) have
shown how these simple assumptions, along with standard
assumptions about the trade-off between decay and rehearsal,
can produce serial position functions and error gradients that
mimic the patterns normally found in short-term serial recall.
In many respects, the primacy model attempts to formalize
the main features of Baddeley’s phonological loop. By adding
specific mechanisms—for example, primacy-based activation
gradients, noisy selection processes, suppression—it becomesFigure 15.4 The multicomponent working memory model proposed by
Baddeley (2000). The central executive controls three slave systems (visuo-
spatial sketchpad; episodic buffer; phonological loop); each interacts with
long-term memory and knowledge to improve on-line performance.