AuthorsGeng Li, Yuhao Wang, Dong Li et al.
TL;DR
GraphMemix builds a query-aware evidence forest over multimodal memories, using anchor-conditioned graph optimization to reach 61.55% Judge Accuracy (macro) vs 49.80% for UniversalRAG (+11.75pp).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing question-agnostic memory systems must process the entire user history to avoid omissions, creating large cold-start and update costs and still dropping decisive visual or local context.
Similarity-based multimodal RAG overselects near-duplicate records and misses low-similarity complementary evidence, even though GraphMemix’s benchmarks show many answers require multiple memories whose joint utility is not captured by similarity alone.
HOW IT WORKS
GraphMemix centers on three components: candidate graph construction, evidence utility and activation costs, and forest optimization, turning multimodal memories into a query-aware evidence forest.
You can think of GraphMemix as a smart card catalog plus wiring: it first finds relevant cards, then checks which cards truly help together, and finally wires them into trusted trees under a context budget.
This anchor-conditioned forest optimization lets GraphMemix recover low-similarity but necessary evidence and suppress redundant or conflicting memories, something a plain context window or top-K similarity ranking cannot do.
DIAGRAM
This diagram shows how GraphMemix processes a single question: from multi-view seed retrieval through graph expansion, verification, and forest selection.
DIAGRAM
This diagram shows how GraphMemix is evaluated across four multimodal memory benchmarks, including lifecycle cost measurement and judge-based accuracy.
PROCESS
Candidate Graph Construction
GraphMemix runs multi-view retrieval over text, images, video, and OCR, then expands top-L seeds via schema and mutual KNN edges into a bounded candidate subgraph C_q.
Evidence Utility and Activation Costs
GraphMemix uses a listwise node verifier to compute p_i(q) and an Evidence-Chain Verifier to assign incremental-support scores, roles, and edge costs c_e(q) for anchor-conditioned schema relations.
Forest Optimization
GraphMemix maximizes node utility minus edge uncertainty and component cost, using a two-stage solver with 1-swap local search and exact maximum-weight forest computation via Kruskal’s algorithm.
Evidence Forest Serialization
GraphMemix orders components by root utility, traverses each tree breadth-first, and serializes at most K multimodal memories into O_q for the frozen LVLM reader to answer the question.
KEY CONTRIBUTIONS
Query-Conditioned Evidence Forest Formulation
GraphMemix formulates long-term multimodal memory selection as a forest optimization objective Unode(S|q) − Cedge(F|q) − Copen(S,F), enabling adaptive cardinality and structured evidence sets under a reader budget K.
Anchor-Conditioned Evidence-Chain Verifier
GraphMemix introduces an Evidence-Chain Verifier that assigns roles like new_fact, clarification, redundant, and conflict, and computes incremental-support scores s_inc_ai(q) to gate schema edges and suppress unreliable expansion.
Lifecycle-Cost Pareto Frontier on Multimodal Benchmarks
GraphMemix achieves 61.55% macro Judge Accuracy with Qwen3-VL and 67.42% with Gemma 4 on ATM, Mem-Gallery, MemEye, and H2HMem, while shortening ATM-Bench’s full lifecycle by up to 4.74× compared to LightMem.
RESULTS
Judge Accuracy macro (Qwen3-VL)
61.55%
+11.75pp over UniversalRAG (49.80%)
ATM Judge Accuracy (Qwen3-VL)
55.27%
+6.80pp over MemGuide (48.47%)
Mem-Gallery Judge Accuracy (Qwen3-VL)
76.33%
+12.57pp over UniversalRAG (63.76%)
H2HMem Judge Accuracy (Gemma 4)
63.47%
+15.14pp over UniversalRAG (48.34%)
On ATM-Bench, Mem-Gallery, MemEye, and H2HMem, GraphMemix is evaluated with Qwen3-VL-8B-Instruct and Gemma 4 12B Unified using GPT-5-mini as judge. These numbers show GraphMemix’s query-local graph organization consistently improves answer correctness over strong multimodal baselines like UniversalRAG and MemGuide.
BENCHMARK
On ATM-Bench, Mem-Gallery, MemEye, and H2HMem, GraphMemix is evaluated with Qwen3-VL-8B-Instruct and Gemma 4 12B Unified using GPT-5-mini as judge. These numbers show GraphMemix’s query-local graph organization consistently improves answer correctness over strong multimodal baselines like UniversalRAG and MemGuide.
BENCHMARK
Judge Accuracy (%) for GraphMemix versus key baselines on ATM, Mem-Gallery, MemEye, and H2HMem (macro-average).
BENCHMARK
Judge Accuracy (%) for GraphMemix versus key baselines on ATM, Mem-Gallery, MemEye, and H2HMem (macro-average).
KEY INSIGHT
GraphMemix’s bounded candidate graph reveals that 37.87% of Mem-Gallery’s gold evidence is recoverable only via relational expansion, not direct top-10 similarity.
This is counterintuitive because many retrieval pipelines assume high-similarity memories suffice, yet GraphMemix shows crucial evidence often lives in low-similarity neighbors reachable only through schema and anchor-conditioned edges.
WHY IT MATTERS
GraphMemix unlocks query-time reconstruction of a structured evidence forest, letting agents reason over long multimodal histories without expensive global summarization.
Builders can now design memory systems that recover complementary, low-similarity evidence and maintain trusted relational context, enabling more reliable long-horizon personal assistants and multimodal RAG agents.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…