Tencent releases Hy-Memory-30B-A3B, a 3B-active model that writes agent memory for Mem0 and MemOS

Tencent Hunyuan has released Hy-Memory-30B-A3B, a 30B-parameter model with roughly 3B active parameters, designed to run the write path inside external long-term-memory frameworks. Weights are on Hugging Face in BF16/F32 safetensors. A technical report (arXiv 2610.22196) and evaluation code (GitHub: Tencent-Hunyuan/Hy-Memory) are published alongside.

What the model does

Hy-Memory-30B-A3B is a drop-in replacement for the general-purpose LLM that Mem0, MemOS, and Tencent's own Hy-Memory framework (Chat, Working, and Coding scenarios) use to write memory. It decides what to retain, how precisely to represent it, and how new evidence should update the persistent memory bank. It does not serve as a general chat model.

Training

According to Tencent, training has two stages, both run inside the same deployed harness configurations:

1. Memory On-Policy Distillation (OPD): dense teacher supervision on state chains the student writes itself.
2. Attributed Memory Reinforcement Learning: the RL stage examines the resulting memory banks, assigns chunk-level credit, and carries the best-scoring bank forward as the parent of the next window.

EverMemOS was held out of training to test transfer to an unseen framework.

Benchmark results (Tencent-reported)

The evaluation varies only the memory writer. Harness version, prompts, retrieval configuration, answer model, and judge are fixed. Every benchmark was run on a fixed 20% subsample. Each harness score averages six metrics: LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench. The overall score averages the four harnesses.

Writer	Mem0	Hy-Memory	MemOS	EverMemOS (held out)	Avg.
Hy-Memory-30B-A3B	54.02	60.55	50.51	56.89	55.49
Claude Opus 5	53.43	60.53	50.50	54.04	54.62
DeepSeek V4.1 Flash	49.73	60.20	49.17	55.47	53.64
Hy3	45.99	58.06	50.19	52.30	51.63
Qwen3.6-35B-A3B	47.51	55.87	47.36	50.75	50.37
Hy3-A3B Base	40.20	50.39	49.05	50.98	47.66
Qwen3-30B-A3B	39.81	45.67	45.44	46.57	44.37

The margins at the top are narrow. Hy-Memory-30B-A3B leads Claude Opus 5 by 0.87 points on average and by 0.02 on the Hy-Memory harness. It does not win every sub-metric. Claude Opus 5 scores 69.47 on LoCoMo in the Mem0 harness versus 59.54, and DeepSeek V4.1 Flash scores 82.83 on LongMemEval in EverMemOS versus 76.00. Shopping PS scores are low for every model, between roughly 14 and 24.

These results come from Tencent's own report and have not been independently reproduced. The comparison also uses a 20% subsample of each benchmark.

Deployment

Tencent documents serving through vLLM (tested with v0.31.0) behind an OpenAI-compatible API. The quickstart enables a no-think mode via reasoning_effort: "no_think" in the chat template kwargs, with temperature 0.0. No Hugging Face Inference Provider currently hosts the model. Hugging Face lists 13 quantizations and one finetune, and 736 downloads in the last month.

Not disclosed in the source material: context window, license terms, training data, training cutoff, and hosted API pricing.

What this means

Memory writing is a high-volume, repeated operation: every conversation chunk triggers a write. Frontier models such as Claude Opus 5 are expensive for that job. A ~3B-active model that Tencent claims matches or beats them within these harnesses would cut the per-write cost of agent memory substantially, if the results hold up.

The held-out EverMemOS result (56.89, the highest in the table) is the most useful signal, since it suggests the training is not simply overfit to the three harnesses it saw. Independent replication on full benchmarks, not 20% subsamples, is the missing check. The gaps are small enough that harness choice and sampling noise could change the ranking.