Individual differences in learning ability of working memory (WM) is known to exist, yet little is known about neural constraints on the ability. Studies on resting-state brain activity provided insights into individual differences in behavior. Previous findings suggested learning ability in cognitive tasks is partly determined by functional connectivity (FC) of task-relevant brain regions. These studies, however, restricted regions of interest to hypothesized areas. Here, we perform whole-brain data-driven analysis to examine if FC patterns predict individual learning plateau. Twenty-nine subjects underwent resting state fMRI scan and received a training of a 3-back task in 90 minutes. A learning plateau was estimated by fitting an inverse curve (y = a - b/x). We calculated FC patterns among 18 networks identified in a massive meta-analysis (the BrainMap ICA). We trained a sparse regressor to identify FC patterns that predict individual plateau. In a leave-one-out cross-validation, the regression model significantly predicted individual plateau (R² = 0.734, p = 0.0025). The sparse regression selected 16.24 ± 2.19 (mean ± SD) out of total 171 connections across validation folds. We counted how many times each connection was selected, and calculated probability of the number of times if 16 were randomly selected from 171 connections (a binominal test). At 9 connections, the probability was less-than 0.05 after Bonferroni correction for 171 connections. Among the 9 connections, the largest contribution (quantified as product of weight in the regression model and correlation in FC) was observed at FC within the left fronto-parietal network, which is most relevant with WM according to metadata of the BrainMap ICA. Positive FC between networks for response selection and execution also contributed to the prediction. Small but significant contributions were observed in multiple negative FCs between WM relevant and less-relevant networks. These results suggest importance of motor networks and efficient suppression of less-relevant networks for high-level WM performance.