GraphMemix: open-source graph memory lifts multimodal agent memory accuracy by up to 11.75 points
TheTuringPost · x · 2026-09-28
MemoraX AI open-sourced GraphMemix (arXiv:2608.26983), reframing long-term memory retrieval for multimodal agents from ranking individual records to building query-aware "evidence forests."
Key idea
- Memories are graph nodes; edges connect memories that add context to each other. The core question: does this memory add something useful beyond what's already found?
- Three components: ① candidate graph construction expanding seed memories via schema and semantic relations; ② evidence utility and activation costs to suppress redundant or conflicting info; ③ forest optimization selecting memory context under a maximum evidence budget.
Results
- Across 4 multimodal memory benchmarks: 61.55% with Qwen3-VL-8B (+11.75 points over the strongest baseline on average) and 67.42% with Gemma 4 12B.
- In one example, it connected clues similarity search missed to correctly identify both dogs: Lumi the Maltese and Coco the Toy Poodle.
- The paper claims a new Pareto frontier between accuracy and lifecycle cost. Paper and code are both open-sourced.
More from coding & agent
- Vibe coding laser-traced visuals: just describe it and let the agent build — natesiggard · 2026-09-28
- Developer turns Mac wallpaper into a living city that animates when agents run — nptacek · 2026-09-28
- 282 viral Claude Opus 5.5 videos with exact prompts, all in one open-source repo — dotey · 2026-09-28
- Dev builds a Grok agent that scans official sources to find and auto-fill money claims you qualify for — elonmusk · 2026-09-28
- Supermemory hits 100k+ weekly NPM installs as go-to memory system for AI agents — julianweisser · 2026-09-28
- Persistent AI instance prefers chatting with its creator; long-context degradation observed — repligate · 2026-09-28