Reproduce “GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Geng Li, Yuhao Wang, Dong Li, Jianye Hao, Yuxin Peng
GraphMemix selects question-relevant memories and verified links from multimodal archives. It retrieves candidates, assesses direct support and whether linked evidence adds information, then chooses a compact forest for answering. Across four benchmarks, it achieves 61.55% macro-average Judge Accuracy with Qwen3-VL-8B-Instruct, versus UniversalRAG’s 49.80%. Judge Accuracy reflects answers judged correct; GraphMemix leads each benchmark in this setting.
Key insights
- Focusing intensive semantic reasoning and optimization on a bounded, query-relevant subgraph drastically reduces memory lifecycle costs compared to global preprocessing approaches.
- Explicitly modeling inter-evidence dependencies and performing joint forest optimization effectively retrieves complementary, low-similarity information that direct retrieval methods typically miss.
- Separating verifier responsibilities for independent node utility and anchor-conditioned relational support enhances accuracy and recall without incurring sequential processing costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.