Store agent episodes in a dedicated event subgraph and use cross-graph traversal for multi-hop temporal questions

arXiv 2604.12285·medium signal

Graph-based agent memory (GAM, MAGMA) keeps episodic events in a separate event subgraph distinct from the atemporal semantic graph, then traverses across both to answer multi-hop 'what happened, when, and why' questions that flat vector logs cannot. The shift is from a passive log of facts to a topological model that preserves how information connects over time. Directly relevant to temporal knowledge-graph builds: separate the 'what is true now' store from the 'what happened when' store rather than collapsing both into one index.