Any extended task episode is subsumed by goal-directed programs that hierar- chically control its execution. We investigated the relationship between working memory capacity and the control instantiated by such hierarchical task entities across four experiments. In a new extended task consisting of subtask A and subtask B, participants first memorized the orientation of subtask A lines (let’s call this event mA), then memorized subtask B lines (mB), then recalled these B lines (rB), and finally recalled A lines (rA). The task structure was: mA-mB-rB- rA. Subtask A lines were thus held in mind during the execution of subtask B. Even though participants had to remember the orientation of lines in both cases, increased WM load of lines A only affected performance on subtask A and did not affect the performance on subtask B. In Experiment 2, four trials of Exp1 were organized into a complex 4-part task with the added condition that A lines of a part be recalled not in that part but in the next part. The task structure was: mA1-mB1-rB1—mA2-mB2-rB2-rA1—mA3-mB3-rB3-rA2—mB3-rB3-rA3. Load of A lines again did not affect B lines. Crucially, load of A2 and A3 lines did not affect the recall of A1 and A2 lines, respectively. In Experiment 3, in a design similar to Exp1, time constraint on mA and mB increased the interference across concurrent subtasks. Experiment 4 showed that increasing the similarity between subtask A and subtask B of Exp1 may increase the across-subtask in terference. We show that WM information of different concurrent subtasks can be maintained separately, perhaps as part of their goal-directed programs. And, encoding to these non-interfering stores, as well as retrieval from them, might depend on attentional and time-based mechanisms.