The　graph　few-shot　class　incremental　learning　(GFSCIL)　task　incrementally　acquires　new　knowledge　from　non-stationary　data　streams　with　limited　training　samples.　It　encompasses　two　distinct　phenomena:　overfitting,　which　occurs　when　the　network　excessively　aligns　with　the　characteristics　of　few-shot　samples　and　fails　to　capture　the　underlying　class　patterns;　and　catastrophic　forgetting,　where　the　introduction　of　new　information　gradually　interferes　with　previously　learned　knowledge　in　the　network.　The　prototype-based　methods　tackle　those　challenges　by　employing　clustering　in　the　metric　space　to　acquire　class　prototypes,　which　differ　from　traditional　class　incremental　learning　methods　in　modifying　model　parameters　or　inserting　regular　term　constraints.　We　propose　the　uncertainty-guided　recurrent　prototype　distillation　network　(URPD)　for　GFSCIL　to　address　the　aforementioned　challenges.　URPD　comprises　two　key　components:　a　recurrent　prototype　representation　(RPR)　module　and　a　generated　distillation　(GD)　module.　The　RPR　module　tackles　the　overfitting　issue　by　generating　recurrent　class　prototypes　based　on　an　uncertainty　selection　scheme　applied　to　unlabeled　nodes.　The　GD　module　mitigates　the　catastrophic　forgetting　issue　by　using　a　generative　distillation　scheme,　which　distills　old　knowledge　based　on　not　only　current　nodes　but　also　generated　replay　nodes.　Essentially,　URPD　improves　traditional　prototype-based　methods　by　learning　debiased　class　prototypes　with　more　fruitful　knowledge　induced　from　unlabeled　nodes.　Experimental　results　demonstrate　that　URPD　outperforms　current　state-of-the-art　methods　by　a　margin　ranging　from　0.95　to　6.46%　on　the　Cora-Full,　Cora-ML,　Flickr,　and　Amazon　datasets.