Hy-Memory
Tencent's 30B MoE (3B active) trained only to write long-term memories for Mem0, MemOS and Hy-Memory. Apache 2.0.
In short. Hy-Memory is a specialist model for the memory side of an AI assistant. After each conversation it decides which facts are worth remembering, how to word them and how to update what is already stored. It is meant to replace the big general-purpose model that memory tools such as Mem0 and MemOS usually call for this job.
Summary
In an external memory system such as Mem0, MemOS or Tencent's own Hy-Memory framework, a language model decides after each exchange what to keep, how precisely to write it down and how new evidence should change what is already stored. Tencent Hunyuan trained Hy-Memory-30B-A3B, a mixture-of-experts model with about 3B parameters active per token, to do that job in place of a general-purpose model. Its Hugging Face repository appeared on 15 October 2026 and the technical report is dated the same day. The config uses the HYV3 architecture class that the Hy3 family also uses, with 128 experts, 8 of them active per token, and a position limit of 262,144 tokens, and the card lists English and Chinese.
Training has two stages. Memory On-Policy Distillation teaches each framework's write protocol on states the student model wrote itself, and Memory RL then optimizes downstream answerability with per-chunk probes, rewards attributed across chunks, over-storage controls and chunk-level GRPO. On fixed 20% subsamples of its benchmarks, Tencent reports a 55.49 macro average over four harnesses, against 54.62 for Claude Opus 5, 53.64 for DeepSeek V4.1 Flash, 51.63 for Hy3 and 50.37 for Qwen3.6-35B-A3B. EverMemOS was held out from training, and Hy-Memory scores 56.89 on it against 54.04 for Claude Opus 5. The margin over Claude Opus 5 is only 0.02 points on the Hy-Memory harness and 0.01 on MemOS.
In the evaluation only the memory writer changes, while the harness version, prompts, retrieval setup, answer model and judge stay fixed. The quickstart serves the model with vLLM, tested with 0.31.0, behind an OpenAI-compatible API and extracts user facts as a JSON list with no-think mode on. The BF16 weights are about 60.1 GB. Tencent also publishes an FP8 build of about 30.6 GB and GGUF builds from 8 GB to 32 GB, and its GGUF card reports 57.25 on one LoCoMo run for the roughly 8 GB expert-mixed build, against 58.78 for its best Q6_K build. The license is Apache 2.0, and the LICENSE file says it adds no restrictions. In the 30 days to 11 November 2026, Hugging Face counted about 7,400 downloads for the BF16 repository and about 34,800 for the official GGUF repository.
Good for
- Replacing the general-purpose model in a Mem0, MemOS or Hy-Memory pipeline with a smaller, self-hosted memory writer
- Cutting the cost of the write path, since only about 3B parameters are active per token
- Trying a memory writer on another framework, because the card says it transfers to unseen ones such as EverMemOS
- Running the writer on a single machine with the official FP8 or Q4_K_M GGUF builds, which are about 31 GB and 18 GB
Watch out for
- It is a specialist for writing memories, not a general chat model, and the card reports only memory benchmarks
- All results come from Tencent on fixed 20% subsamples, and the lead over Claude Opus 5 is only 0.02 and 0.01 points on two of the four harnesses
- It was trained inside specific harness configurations of Mem0, MemOS and Hy-Memory, so check results on your own framework version and prompts
- The GGUF scores are single runs on one LoCoMo setup, and Tencent says they do not claim lossless quantization
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Hy-Memory-30B-A3B (Q4_K_M GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Hy-Memory-30B-A3B BF16 | 30.1B 3B active | 69 GB | 35 GB | 21 GB | Gaming GPU |
| Hy-Memory-30B-A3B-FP8 FP8 | 30.1B 3B active | 69 GB | 35 GB | 21 GB | Gaming GPU |
| Hy-Memory-30B-A3B (Q4_K_M GGUF) INT4 | 30.1B 3B active | 69 GB | 35 GB | 20 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is Hy-Memory?
- Hy-Memory is a specialist model for the memory side of an AI assistant. After each conversation it decides which facts are worth remembering, how to word them and how to update what is already stored. It is meant to replace the big general-purpose model that memory tools such as Mem0 and MemOS usually call for this job.
- Can I use Hy-Memory commercially?
- Yes. Hy-Memory is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Hy-Memory open source?
- Hy-Memory is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Hy-Memory on my own computer?
- Yes, on a good graphics card. The smallest variant, Hy-Memory-30B-A3B (Q4_K_M GGUF), needs about 20 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Hy-Memory released?
- Hy-Memory was released by Tencent on October 15, 2026.
- How long a context window does Hy-Memory have?
- Hy-Memory supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing (arXiv, 15 Oct 2026) arxiv.org
- Hy-Memory-30B-A3B model card (Hugging Face) huggingface.co
- Hy-Memory-30B-A3B-GGUF model card (Hugging Face) huggingface.co
- Hy-Memory repository (GitHub) github.com
- Hy-Memory-30B-A3B LICENSE (Apache 2.0) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MemT-1.0 MemTensor | 35.1B – 288B 17B active | 256K | Apache 2.0 (Permissive) | 24 GB Gaming GPU | |
| Hy3 Tencent | 295B 21B active | 256K | Apache 2.0 (Permissive) | 204 GB Big-memory desktop | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server |