Two-phase model of working memory (WM) search tries to explain memory search taking set size, positive vs negative trials difference and speed of stimuli presentation into consideration. It was assumed that in the first - parallel - phase cognitive system uses both bottom-up (items in memory) and top-down (target) activations to prepare to actual search. Preparation to search is a process of indexing memory contents that uses strength of items activation or their similarity to target. The second phase is a serial process of comparing the target with items stored in memory which leads to signal detection (positive trials) or recognition of its absence (negative trials). This paper describes an implementation of the model for verbal stimuli in ACT-R cognitive architecture. Two-phase model has been extended with two strategies of memory trace coding. Computational model simulates 95,4% and 97,4% variance observed in accuracy and reaction latencies, respectively. It replicates group differences in WM scanning. A comparison of alternative model versions falsifies single-phase models of WM scanning. A third search phase was needed to obtain best fit to the data: parallel and non-efficient search of partially active part of WM that is outside the focus of attention.