Tencent's Hy-Memory-30B-A3B: A small model that writes long-term memory for AI

What this is about

Tencent Hunyuan has released Hy-Memory-30B-A3B on Hugging Face, a model under the Apache 2.0 license. It isn't a general-purpose chatbot. It's a specialist for the long-term memory of AI systems. Such memory systems store information from conversations outside the model, in a kind of notebook (memory bank), so an assistant can later recall preferences or facts. The model takes over the "write path": it decides which information to keep, how exactly to word it, and how new clues change the existing memory.

Hy-Memory has 30.1B parameters (roughly speaking, the tunable values learned during training), of which only about 3B are active per token. That's the Mixture-of-Experts design, where only some of the sub-networks compute at any one time. It saves compute. The model understands English and Chinese.

What it's for

According to Tencent, the model is a drop-in replacement for the general-purpose language model in the Mem0, MemOS and Hy-Memory frameworks (with the Chat, Working and Coding scenarios). It also transfers to an unseen framework: EverMemOS was not part of the training. In the model card's example, it extracts explicit facts and preferences from a conversation (such as name, city, programming language, vegetarian diet) as short, self-contained sentences in JSON format, without inventing anything. The "no_think" mode switches off longer reasoning.

Two-stage training

Both training stages run in the same frameworks the model is later used in:

1. Memory On-Policy Distillation (OPD): A teacher model gives dense feedback on memory states that the student model wrote itself. Distillation means a smaller model learns from a larger one.
2. Attributed Memory Reinforcement Learning: In reinforcement learning, the resulting memory stores are scored, and the reward is attributed to individual text segments (chunks). The best memory state from one run is passed on as the starting point for the next window.

Results (developer-reported)

Only the memory writer was tested. Framework version, prompts, retrieval, answer model and judge model were held fixed. To save compute, Tencent used fixed 20-percent samples of each benchmark. Six scores are averaged per framework: LoCoMo, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and Code Memory Bench.

Hy-Memory leads on average, though sometimes only narrowly: on the Hy-Memory framework and MemOS, the gap to Claude Opus 5 is about 0.02 and 0.01 points. The individual tables show a mixed picture. Opus 5 is clearly better on LoCoMo with Mem0 (69.47 versus 59.54), for example, while Hy-Memory wins on LongMemEval with Mem0 (73.00 versus 66.00) and Travel SPS (54.51 versus 42.98). Against its base model Hy3-A3B Base, the average rises from 47.66 to 55.49, and against Qwen3-30B-A3B from 44.37 to 55.49. The jump on the unseen EverMemOS in Code Memory Bench stands out (63.19 versus 49.53 for Opus 5).

What you need to try it

The source gives no hardware figures. As an estimate, a 30B model needs about 18–20 GB of VRAM at 4-bit quantization and around 60 GB at full precision. The source lists no GGUF version. The quickstart runs on vLLM (tested with v0.31.0) with --served-model-name hy-memory and an OpenAI-compatible API. Plan for about 60 GB of disk space for the weights. Without a GGUF conversion, running it on a Mac or CPU is impractical.

Our take

Memory is a core problem for personal AI assistants. A small, open specialist model that can replace expensive API models as the writer is practically interesting, and just 3B active parameters make always-on operation cheaper. Still, caution is warranted. All results come from Tencent itself, on 20-percent samples, with small margins that could be noise at such sample sizes. Only the write path is tested. Quality in real long-running use, privacy for stored personal facts, and robustness outside the tested frameworks all remain open questions. If you run Mem0, MemOS or similar systems, it's worth comparing it in your own scenario.