Dynamic　graph　representation　learning　has　caused　much　attention　in　many　practical　applications.　There　is　an　interesting　method　that　uses　RNNS　(e.g.,　LSTM,　GRU)　to　update　the　GCN’s　weights　dynamically　with　weights　from　the　previous　time　step.　However,　it　only　uses　one　time-step　weights,　which　eventually　leads　to　the　lack　of　sufficient　historical　information.　In　this　work,　we　focus　on　this　method　for　the　developing　parameters　and　propose　a　developing　GCN　model,　which　adapts　an　attention　mechanism　to　get　richer　historical　information　so　that　the　RNNs　will　decode　better　fused　historical　representations　to　capture　the　temporal　correlation　of　weights　in　the　GCN,　which　not　only　can　learn　those　dynamic　graphs　with　fewer　features,　but　also　can　extract　richer　historical　information　to　learn.　We　evaluate　our　method　on　the　task　for　link　prediction　and　the　result　shows　a　better　performance　in　most　data　sets　we　test.　©　2021,　Springer　Nature　Switzerland　AG.