Betty's Brain is an open-ended learning environment in which students learn about science topics by teaching a virtual agent named Betty through the construction of a visual causal map that represents the relevant science phenomena. The task is complex, and success requires the use of metacognitive strategies that support knowledge acquisition, causal map construction, and progress monitoring. Previous research has established that middle school students struggle at such tasks without proper scaffolding and feedback. In Betty's Brain, this feedback is provided by Betty and Mr. Davis, another virtual agent designed to provide guidance and suggestions as students work. This paper discusses our implementation of contextualized conversational (CC) feedback, and then presents the results of an experimental study exploring the effects of this feedback in two 8th-grade science classrooms. The results illustrate some advantages of the CC feedback in comparison with a baseline dialogue mechanism that presents similar strategies in a non-conversational, non-contextualized form. While both groups showed significant pre-to-post test learning gains, the difference in learning gains between the groups was not statistically significant. However, students who received CC feedback more often performed actions in accordance with the advised strategies, and they created higher quality causal maps.