Tencent Hunyuan open-sources Hy Memory 2.0 memory framework and model
Tencent Hunyuan released Hy Memory 2.0, an open-source stack consisting of the Hy-Memory-30B-A3B memory model (roughly 30B total / ~3B active parameters, Apache-2.0), a memory framework SDK, and MemHub, a local memory app with a CLI and Mac client. The model is trained specifically for the write path inside external memory frameworks, replacing the general-purpose LLM in Mem0, MemOS and Hy-Memory by deciding what to retain and how new evidence updates the memory bank; the company says it ranks first on four benchmarks spanning chat, work and code memory, with work and code memory newly added in 2.0. Per the company and its arXiv paper, across the Mem0, Hy-Memory, MemOS and held-out EverMemOS harnesses on a fixed 20% benchmark subsample, the model scored 55.49% macro-average, above Claude Opus 5 at 54.62%, DeepSeek V4.1 Flash at 53.64% and Qwen3-30B-A3B at 44.37%. IT之家 also reports chat memory accuracy of 72.6% on LoCoMo with per-question tokens falling from 1,159 to 703, and says MemHub plugs into 10 AI tools including WorkBuddy, Cursor, Codex, Qwen Code, DeepSeek Harness and OpenCode.
Memory writing is a bottleneck for long-horizon agents, so a small ~3B-active specialist replacing a general model for memory management could cut token cost while improving continuity across sessions and tools.