It has long been recognized that our ability to actively attend to and concurrently process information is limited (Broadbent, 1958). Nonetheless, component operations in many cognitive skills often rely on the products of prior perceptual and cognitive analyses. For example, subgoals in problem solving and reasoning can rely on the products of earlier subgoals (Anderson, 1983). Similarly, in language understanding, comprehenders frequently need to resolve dependencies between elements separated by several phrases or clauses (McElree, 2000; McElree, Foraker, & Dyer, 2003). As ongoing operations will often displace past analyses from the current focus of attention, successful execution of many operations may depend on our ability to rapidly shunt information between memory and focal attention. Many researchers have suggested that a working memory (WM) system partially compensates for our limited capacity to concurrently process information. Working memory is thought to provide a "workspace," where a few by-products of recent perceptual and cognitive processing can be maintained in a more accessible state than information in long-term memory (LTM). Working memory representations may be more accessible than LTM representations either because they are held in specialized stores (Baddeley, 1986; Baddeley & Hitch, 1974; Schneider & Detweiler, 1988; Shallice & Vallar, 1990) or simply because they have residual activation from recent processing (Anderson, 1983; Conway & Engle, 1994; Cowan, 1995, 2001; Engle, 1996; Ericsson & Pennington, 1993). If one posits a distinct WM system, as illustrated in Fig. 1A, information can be represented in three possible states, either in LTM, WM, or in the current focus of attention. Different forms of evidence have been used to motivate this type of tripartite architecture (Cowan, 1995, 2001). However, the evidence is indirect and can be challenged on several grounds (Crowder, 1993; Nairne, 1996; Wickelgren, 1973). This chapter examines whether there is direct evidence for either qualitative or quantitative differences in retrieval for items that the framework in Fig. 1A posits to be in distinct representational states. Different states could be motivated by findings that a qualitatively different type of retrieval operation is used to access information in each state. Alternatively or additionally, the architecture in Fig. 1A could be motivated by discontinuities in retrieval speed. This prediction is illustrated in Fig. 1B. The ultimate success of retrieval will be limited by forgetting due to the passage of time or intervening items between study and test, which should lead to systematic declines in accuracy with diminished recency. However, a straightforward prediction of a tripartite architecture is that each state should be associated with a distinct retrieval speed. Information in the current focus of attention should exhibit a privileged form of access. Less recent representations-those that are outside the capacity of focal attention but still within the span of WM-should be accessed slower than items within focal attention but faster than LTM representations. Finally, information that resides in LTM should be associated with the slowest retrieval speed. This chapter reviews studies on the speed and accuracy of accessing representations of recently processed information. It documents the types of retrieval operations used to access both item and order information. Evidence is presented indicating that item information is retrieved with a direct-access (content-addressable) process (Section II. B), whereas order information is retrieved by a slower serial search process (Section II. C). Crucially, however, in neither case do we find evidence for a qualitative or quantitative "break-point" between what a tripartite architecture posits as the divide between WM and LTM. Collectively, the temporal dynamics of retrieval are indicative of two rather than three representational states. These measures provide clear evidence for a distinction between information within the current focus of attention and information passively stored in memory but not a further distinction corresponding to WM and LTM. Rather, the evidence suggests the type of dichotomy illustrated in Fig. 2A in which there is only an architectural difference between representations in focal attention and representations in memory. The corresponding speed and accuracy profiles are illustrated in Fig. 2B. Like Fig. 1B, the probability of retrieving an item from memory decreases continuously as more information is interpolated between study and test. Here, however, accessibility shows a sharply dichotomous pattern. Items within focal attention are accessed quickly, but all other items outside attention are accessed more slowly and with the same retrieval speed. The architecture in Fig. 2A hearkens back to the simple dichotomy James (1890) drew between primary memory, which he regarded as synonymous with conscious awareness, and secondary memory, the repository of all passive memory representations. The second emphasis of the chapter is on the capacity of focal attention. Information in focal attention can be discriminated from information in a more passive state by its relatively fast retrieval dynamics (Section III. A). This view is reinforced by two new experiments that measure changes in retrieval dynamics that result from explicit attempts to shunt information from memory to focal attention (Sections III. B and III. C). Estimates of focal attention based on retrieval speed measures indicate that it has a much smaller capacity than has typically been assumed in some current approaches (Cowan, 2001). This view is reinforced further by studies that challenge subjects to attempt to retain items in focal attention while concurrently processing other information (Section III. D). The chapter ends with a brief discussion of neuroimaging findings.