Motor　imagery　electroencephalogram　(MI-EEG)　plays　an　important　role　in　the　brain　computer　interface-based　neuro-rehabilitation,　and　it　is　challenging　to　make　the　recognition　model　have　the　self-learning　and　updating　ability　with　the　recovery　process.　Broad　learning　system　(BLS)　may　be　perfect　duo　to　the　incremental　learning　and　dynamic　broad　expansion.　However,　its　natural　shallow　network　and　simple　mappings　in　generating　feature　nodes　and　enhancement　nodes　limit　the　feature　mapping　capacity　and　yield　performance　degradation　for　new　data.　In　this　paper,　we　develop　a　session-incremental　broad　learning　system　(SIBLS)　to　adaptively　recognize　the　incremental　sessions　of　MI-EEG　with　changing　probability　distributions.　A　multiscale　feature　mapping　block　(MFMB)　is　designed　to　extract　the　global　and　local　features　for　reinforcing　feature　mapping.　Then,　the　multiple　key　feature　distillation　blocks　(KFDB),　which　are　constructed　by　manifold　learning　embedded　sparse　autoencoders,　are　further　employed　to　extract　compact　and　highly　geometrically　relevant　features　for　producing　enhancement　nodes.　For　the　incremental　sessions,　Pearson＇s　correlation　coefficient　constraints　(PCCs)　and　taskdependent　correlation　constraints　(TCs)　are　added　to　KFDB　and　output　layer　respectively　to　make　the　same　category　closer　between　adjacent　sessions.　The　experiments　are　conducted　on　two　public　datasets,　which　include　two　and　five　sessions　respectively,　the　time-frequency　spectrum　of　each　electrode　is　simultaneously　input　to　SIBLS,　separately　achieving　the　average　accuracies　of　all　sessions　79.62%　and　87.59%.　Experiment　results　show　that　the　perfect　plasticity　of　SIBLS　derives　in　the　main　from　the　strong　feature　mapping　of　MFMB　and　KFDB,　and　the　forgetting　is　effectively　inhibited　by　introducing　constraints　as　well.