Long-term multimodal agent memory as a query-aware evidence forest instead of offline summaries or embedding lookup

arXiv 2608.26983·low signalXBlueskyLinkedInCopy link

GraphMemix treats memory organization as combinatorial optimization: expand multi-view seed memories through schema and semantic relations, decouple direct evidence support from anchor-conditioned relation verification to suppress redundant or conflicting context, then jointly select a forest-shaped context under a maximum evidence budget. The point is avoiding both the lifecycle cost of question-agnostic offline summarization and the incompleteness of naive similarity matching, while recovering low-similarity complementary evidence that top-k retrieval drops. It sets a new accuracy/lifecycle-cost Pareto frontier on four long-term multimodal memory benchmarks.