GraphMemix is a new framework for organizing long-term memory in multimodal agents. It treats memory selection as a combinatorial optimization problem: given a query, it builds a candidate graph of memories and then selects a forest (a set of nodes and edges) that maximizes evidence utility under a budget. Node and edge verifiers assess direct support and complementary value, while Kruskal's algorithm finds the optimal forest exactly.
Experiments on four multimodal memory benchmarks show that GraphMemix improves judge accuracy by 11.75 percentage points (from 49.80% to 61.55% macro-average) over the best public baseline using Qwen3-VL. It also achieves a better accuracy versus lifecycle cost tradeoff, establishing a new Pareto frontier. This is interesting because it avoids the high cost of offline summarization and overcomes the limitations of similarity-based retrieval, which often misses complementary low-similarity memories.