Organizing long‑term memory for multimodal agents remains a challenging problem. Existing approaches either rely on expensive, query‑agnostic offline summaries or on naive embedding‑similarity matching, which yields incomplete and redundant context. To tackle these issues we introduce GraphMemix, a combinatorial‑optimization graph‑memory framework that treats memory organization as a query‑aware evidence‑forest construction.
The method consists of three key components:
- Candidate graph construction – expands multi‑view seed memories via schema and semantic relations to obtain query‑relevant original context.
- Evidence utility and activation cost – decouples direct memory support from anchor‑conditioned relation verification, suppressing redundant or conflicting information.
- Forest optimization – jointly selects a forest‑structured memory subgraph under a maximum evidence budget while preserving a reliable relational structure.
By structuring memory as a query‑relevant subgraph, the approach avoids substantial lifecycle overhead and recovers low‑similarity complementary evidence. Experiments on four long‑term multimodal memory benchmarks show significant gains across various foundation models and establish a new Pareto frontier between accuracy and lifecycle cost.
Blogger's Review: GraphMemix achieves impressive accuracy while markedly reducing retrieval cost, highlighting the power of query‑aware memory organization.