Tencent Hunyuan released Hy-Memory-30B-A3B, a compact long-term-memory writer model
Tencent Hunyuan released Hy-Memory-30B-A3B, a ~3B active parameter long-term-memory writer model, on Hugging Face under Apache-2.0. It replaces general-purpose LLMs in Mem0, MemOS, and Hy-Memory frameworks. The model is trained via memory on-policy distillation and attributed memory reinforcement learning.
China context
- Original name
- 腾讯混元
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can integrate Hy-Memory-30B-A3B into existing memory frameworks like Mem0 and MemOS to improve memory write quality without changing other components, potentially reducing costs compared to using larger general-purpose LLMs.
- For investors
- The release of a specialized open-weight memory model by Tencent signals competition in the memory layer of AI applications, which could affect the value proposition of companies offering memory solutions or relying on general-purpose LLMs for memory tasks.
Hy-Memory-30B-A3B is a compact long-term-memory writer model developed by Tencent Hunyuan to execute the write path inside external memory frameworks. It directly replaces the general-purpose LLM inside Mem0, MemOS, and Hy-Memory, deciding what information to retain, how precisely to represent it, and how new evidence should update the persistent memory bank. The model is trained in two stages—Memory On-Policy Distillation (OPD) followed by Attributed Memory Reinforcement Learning—and transfers directly to unseen memory frameworks such as EverMemOS.
The model uses a two-stage training process: Memory On-Policy Distillation provides dense teacher supervision on student-written state chains, while Attributed Memory Reinforcement Learning examines resulting banks, assigns chunk-level credit, and carries the selected bank across windows. Evaluation changes only the memory writer, keeping harness version, prompts, retrieval configuration, answer model, judge, and other framework-side components fixed. Cross-harness macro-average scores on a fixed 20% benchmark matrix show Hy-Memory-30B-A3B achieving 55.49 overall, outperforming Claude Opus 5 (54.62), DeepSeek V4.1 Flash (53.64), and other baselines.
Developers using Mem0, MemOS, or Hy-Memory can swap in Hy-Memory-30B-A3B to improve memory write quality without changing their existing retrieval or answer models, potentially reducing reliance on larger general-purpose LLMs for memory tasks. The model's Apache-2.0 license and Hugging Face availability lower the barrier for integration and experimentation.
Hy-Memory-30B-A3B offers a specialized, open-weight alternative for memory management in AI applications, potentially reducing costs and improving performance for developers building with memory frameworks. Its Apache-2.0 license allows commercial use, which could accelerate adoption in enterprise and consumer AI products.
The model's transfer to the held-out EverMemOS harness suggests potential for broader adoption across memory frameworks. Verification of real-world performance outside the benchmark matrix and adoption by third-party developers will be key signals to watch.