Short-Term or Working Memory 433presentation of a single probe; this probe either matches, or
not, one of the just-presented items (e.g., Sternberg, 1966).
Obviously, a task of this kind is relatively easy, and people
rarely make mistakes. Of main interest is the latency, or reac-
tion time, of correct responses, usually as a function of a vari-
able like list length.
Sternberg found that mean reaction time increased linearly
with list length, but the slopes of the reaction time functions
were roughly equivalent for positive and negative responses.
This suggested that people search short-term memory in a
serial, or item-by-item, fashion looking for a match to the
recognition probe; the more items that need to be searched,
the longer the reaction times. Equal positive and negative
slopes suggested as well that the search process was exhaus-
tive, meaning that all the items in the set were compared
regardless of when (or whether) a match was found. The
proposed serial exhaustive search process seemed to rule
out other plausible search procedures—for example, self-
terminating (in which the search process stops once a match
is found) or parallel processing (in which all activated items
are compared simultaneously with the probe).
However, it turns out that mean reaction time cannot be
used to discriminate definitively between serial and parallel
search processes; it is possible to mimic the reaction time pat-
terns noted by Sternberg (1966), for example, using a parallel
processing mechanism with certain additional assumptions
(e.g., Townsend & Ashby, 1983). More diagnostic evidence
comes from either a fuller analysis of reaction time distribu-
tions (e.g., Ratcliff, 1978; Reed, 1976) or from techniques
that examine the full time course of processing during recog-
nition decisions. In the latter instance, the response-signal
procedure cues the subject to respond at particular times after
the appearance of the recognition probe: For example, the
cue to respond might appear almost immediately after ap-
pearance of the probe, which yields performance near chance
levels, or seconds later, which allows for the most accurate
performance. One can then determine how accuracy unfolds
over time—so-called retrieval dynamics—which allows for a
more sensitive analysis of possible retrieval mechanisms (see
Dosher & McElree, 1992; McElree & Dosher, 1989).
Application of the response-signal technique to the re-
trieval of short-term memories in the Sternberg task supports
a parallel, direct-access retrieval process. Retrieval dynamics
seem not to vary much with list length, or serial position,
which is consistent with a parallel matching process (Ratcliff,
1978). The nature of the retrieval process may change, how-
ever, depending on the type of information that must be re-
trieved. McElree and Dosher (1993) report that the recovery
of order information—which of two list items occurred more
recently?—is accomplished through a slow serial retrieval
process; again, this conclusion is based on the finding that the
retrieval dynamics for the order judgment differ systemati-
cally from those found in item recognition (i.e., the Sternberg
task). The fact that the retrieval dynamics vary in this way
undermines the simple notion that activity traces exist in a
state of immediate availability. Even if items exist in a special
focus of attention (e.g., Cowan, 1995), by virtue of their acti-
vation, various kinds of retrieval-based selection processes
are clearly needed to satisfy the demands of differing tasks.
It is possible, however, that there is something special
about retrieval of the very last item, or item chunk, in a short
list. The last item is recognized faster than other items, but
more importantly, the retrieval dynamics appear different as
well (McElree & Dosher, 1989; Wickelgren, Corbett, &
Dosher, 1980). This finding has been interpreted to mean that
the last item remains active in awareness and, thus, can be
matched directly with the recognition probe; the item essen-
tially remains in consciousness, eliminating the need for a re-
trieval mechanism to move it from a passive to an active
state. McElree (1998) recently showed that up to three
items at the end of the list can show these special properties,
as long as they are members of the same category (forming,
presumably, a category chunk). However, alternative inter-
pretations of the data pattern are possible. For example, one
could argue that the contextual cues available at the point of
probe presentation especially match the cues associated with
the last list item; this, in turn, could affect the ease and qual-
ity of the retrieval process.Cue-Driven RetentionOne of the troubling features of direct access (or cueless
retrieval) is the idea that items can be remembered with-
out considering the nature of the retrieval environment. As
Tulving (1983) has argued, there is no justification for mak-
ing absolute statements about the memorability of items—for
example, based on their inherent characteristics or encoding
properties—because remembering always depends on an
interaction between encoding conditions and retrieval con-
ditions. It is possible that short-term memories represent a
special case, violating Tulving’s dictum, but the available
evidence suggests otherwise.
Direct access received some early support from studies
showing that immediate retention, tested without a distractor
interval, can show immunity to proactive interference
(Halford, Mayberry, & Bain, 1988; Tehan & Humphreys,
1995; Wickens et al., 1981). In the Tehan and Humphreys
(1995) experiments, trials consisted of the presentation of
either one or two short four-item lists; the subject’s task was
to recall the last presented list, so on two-list trials subjects
were told to ignore the first list. Of main interest was the
effect of the first list on recall of the second list—that is,