Networked　multi-agent　reinforcement　learning　(MARL)　addresses　distributed　control　of　multi-agent　systems　(MASs).　Traditional　networked　MARL　algorithms　are　unable　to　represent　the　spatial　correlation　of　node　features,　which　results　in　failure　to　guarantee　the　cooperation　and　convergence　performance　of　algorithms.　To　represent　the　spatial　correlation　between　agents,　we　design　an　entity　graph　with　spatial　correlation　for　networked　MARL　(EGNMARL).　Specifically,　each　agent　constructs　a　directed　entity　graph　network　based　on　node　features　from　the　other　entities　relative　to　itself,　and　uses　Euclidean　distances　between　entities　to　represent　edge　features,　which　helps　agents　effectively　understand　their　positions　and　roles　in　MASs.　Second,　we　adopt　the　unified　message　passing　model　(UniMP)　as　the　information　aggregation　strategy　and　embed　it　into　the　actor-critic　learning　framework,　which　enables　the　agents　to　learn　the　node　message　representations　of　their　dynamic　neighbors　and　capture　the　spatial　correlations　among　entities.　Finally,　we　compare　EGNMARL　with　several　baseline　MARL　algorithms　in　different　environments,　and　the　experimental　results　show　that　EGNMARL　achieves　advantages　in　both　cooperation　performance　and　convergence　speed.　©　2024　Copyright　held　by　the　owner/author(s).