Tencent Hunyuan said on October 16 that Hy-Memory 2.0 is open source, releasing a three-part agent-memory stack: the Hy-Memory-30B-A3B writer model, the Hy-Memory framework for storage, updates, and recall, and the MemHub app layer.
Tencent says the system now covers chat memory as well as working-memory and coding-memory scenarios. It also says Hy-Memory supports private deployment and can work with existing LLMs rather than requiring a full model swap.
On performance, Tencent’s model card reports a 55.49 macro average across four memory benchmarks, ahead of Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64 on fixed 20% subsamples. Tencent also says Hy-Memory ranked first among open frameworks on LoCoMo, MemoryArena, CodeMemoryBench, and SWE-ContextBench. Those results are vendor-reported, so the claimed lead still needs broader outside validation.
Why it matters: agent teams now have an open memory model, framework, and app layer to test persistent user, task, and code memory without committing to a single model stack.