Key Info
Tencent open-sourced a full memory stack consisting of the Hy-Memory-30B-A3B model (3B active parameters), the Hy Memory engine, and the MemHub app, claiming it outperforms Opus 5 on memory tasks.
Highlights
- Hy Memory is said to rank first on all four benchmarks run, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%.
- The model is engine-agnostic: dropping Hy-Memory-30B-A3B into Mem0, MemOS, or EverMemOS still comes out on top, even on a system it never trained on.
- It is already deployed in Tencent's own products, where one assistant's multi-hop recall improved from 50% to 90%.
- It unifies memory across Codex, Cursor, WorkBuddy and more; 24GB+ VRAM recommended.
- All three layers are open: MemHub (app), Hy Memory (engine/SDK), and Hy-Memory-30B-A3B (model), with repo and weights released.