Active learning of finite automata has been vigorously pursued for the purposes of analysis and explanation of black-box systems. In this paper, we study an L*-style learning algorithm for weighted automata over the max-plus semiring. The max-plus setting exposes a "consistency" issue in the previously studied semiring-generic extension of L*: we show that it can fail to maintain consistency of tables, and can thus make equivalence queries on obviously wrong hypothesis automata. We present a theoretical fix by a mathematically clean notion of column-closedness. We also present a nontrivial and reasonably broad class of weighted languages over the max-plus semiring in which our algorithm terminates.