Localize the subgraph before you rewrite agent memory: hierarchical graph memory fixes both token cost and conflicting facts

arXiv 2608.05095·medium signalXBlueskyLinkedInCopy link

HiGram replaces flat graph memory with a coarse-to-fine hierarchy so retrieval stops dragging in irrelevant context, uses MicroGraph-based path localization to identify the support subgraph before any update, and rewrites both individual memory units and the relationships between them together. On long-term conversational QA and conflict-aware memory tasks it reports substantial gains in answer quality and token efficiency over baselines, including under dynamic, static, and conditional conflicts. The transferable idea is the localize-then-rewrite order — updating a memory unit without repairing its edges is what produces contradictory recall.