Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing
Tianxiang Fei and colleagues at Tencent Hunyuan train Hy-Memory-30B-A3B, a 3B-active MoE model that handles only the write path of long-term memory (what to store, how precisely, and how to update it), and run it inside three existing memory frameworks.
Ask this paper
One writer, several protocols. The model is trained across Mem0, MemOS and Hy-Memory, so one model can replace the general-purpose LLM that each framework calls through its own prompts.
Two training stages. Memory on-policy distillation teaches each framework's write protocol on states the student generated. Memory RL then optimizes downstream answerability with dense per-chunk probes, rewards attributed across chunks, over-storage penalties and chunk-level GRPO.
Training matches deployment. Memory banks written by the policy are carried across windows during training, so the model learns on the state it will see in use.
Result. On fixed 20% benchmark subsamples it ranks first on all four evaluated harnesses, including held-out EverMemOS, with a 55.49% macro average against 54.62% for Claude Opus 5.
Caveat. The margin over Opus 5 is under one point and measured on subsamples. The stronger claim is cost: a 3B-active writer matches a frontier model on this task.
Abstract
External-memory systems extend language-model applications beyond a context window, but their quality depends critically on the model executing the write path: deciding what to retain, how precisely to represent it, and how new evidence should update persistent state. Existing frameworks typically place a general-purpose large language model behind framework-specific prompts. This raises serving cost, and a writer specialized for one protocol does not transfer directly to another; a unified model that can operate across established memory frameworks remains largely unexplored. We present Hy-Memory-30B-A3B, a compact long-term-memory writer trained across Mem0, MemOS, and Hy-Memory. Memory On-Policy Distillation teaches framework-specific write protocols on student-generated states; Memory RL then optimizes downstream answerability using dense per-chunk probes, attributed cross-chunk rewards, over-storage controls, and chunk-level GRPO. Carrying policy-written banks across windows keeps training aligned with deployment. On fixed 20% benchmark subsamples, Hy-Memory-30B-A3B ranks first on all four evaluated harnesses, including held-out EverMemOS, with a 55.49% macro average versus 54.62% for Claude Opus 5.