Long-term memory for agents needs a model to decide what information is worth keeping, at what granularity, and how new evidence should update old entries. That write path has usually been handed to the main general-purpose LLM — capable, but expensive and overqualified. Tencent Hunyuan's newly open-sourced Hy-Memory-30B-A3B takes a different route: it turns the "memory writer" into a specialist, a 30B-total-parameter MoE that activates only about 3B per token, under an Apache 2.0 license with weights on Hugging Face.

What it does inside memory frameworks

The write path of external memory frameworks — Mem0, MemOS, and Tencent's own Hy-Memory framework covering Chat, Working, and Coding scenarios — used to be staffed by a general-purpose LLM. Hy-Memory-30B-A3B is a direct replacement for that slot: it decides what to retain, how precisely to represent it, and how new evidence updates the persistent memory bank. On the engineering side, the model card notes deployment is tested with vLLM v0.31.0 behind an OpenAI-compatible API.

Two-stage training

Stage one is Memory On-Policy Distillation: the student writes memory inside the same deployed harness configurations, while the teacher supplies dense supervision on the student-written state chains. Stage two is Attributed Memory Reinforcement Learning: the resulting memory banks are examined directly, credit is assigned at the chunk level, and the best bank is carried across windows.

Official eval: the small specialist edges out the flagship

The evaluation changes only the memory writer — harness version, prompts, retrieval configuration, answer model, and judge are all held fixed, and every writer builds its bank from an empty state. Each of the four harnesses (Mem0, Hy-Memory, MemOS, EverMemOS) is scored on an equal-weight average of six metrics — LoCoMo, LongMemEval, PersonaMem-v2, Travel, Shopping, and Code Memory Bench — with the main matrix evaluated on a fixed 20% subsample.

Hy-Memory-30B-A3B posts a four-harness macro average of 55.49, ahead of Claude Opus 5 at 54.62, DeepSeek V4.1 Flash at 53.64, Tencent's own Hy3 at 51.63, and the larger Qwen3.6-35B-A3B at 50.37. The stronger evidence is EverMemOS, held out from training: 56.89 versus 54.04, including a wide 63.19-to-49.53 lead on the code-memory benchmark.

Caveats

This is a vendor-reported evaluation on 20% subsamples, with no third-party replication yet. Nor is it a blanket win: on LoCoMo under the Mem0 harness, Opus 5 scores 69.47 against this model's 59.54. And every writer sits in a low band of roughly 14-24 on the Shopping metric, suggesting preference-style memory is hard across the board.

So what

The takeaway is not "3B beats a flagship" clickbait; it is a validated engineering judgment — memory writing is a narrow, deep task where a small specialist beats a general-purpose giant on cost and latency alike. The tech report is at arXiv:2610.22196 and the code lives in Tencent-Hunyuan/Hy-Memory on GitHub. For teams building agent memory systems, the model card is a ticket to a cheap swap-in experiment.