This study tested the dual-component model of working memory (WM) against its unitary alternative. The former account predicts that WM consists of two functionally distinct mechanisms: a very accessible but capacity-limited primary memory (PM) and a less accessible secondary memory (SM). The latter account assumes only one long-term memory component. We used a novel version of the Sternberg serial recognition paradigm, which selectively impedes access to either early or late items, by asking participants about the location of a probe in relation to either the end or the start of encoded memory set, respectively. When locations matched probes, our manipulation harmed recognition of early items, while it left late items intact, in the case of both latency and accuracy. However, in trials in which locations did not match probes, such an effect regarded only latency but not accuracy. This result suggests that a way of access to WM may depend on the level of conflict among accessed memory items. Finally, confirmatory factor analysis (CFA) revealed two distinct sources of variance in recognition accuracy. In total, our results are consistent with th e dual-component view of WM, and they implicate that early items were presumably held in SM, while late items benefited from being held in PM.