Tencent's Hy-Memory-30B-A3B: A Small Model Built Just to Write Memory
Tencent's Hy-Memory-30B-A3B beats Claude Opus 5 and DeepSeek V4.1 Flash on memory-writing benchmarks. Here's how it works and why it matters.
What is Hy-Memory-30B-A3B?
Hy-Memory-30B-A3B is a specialized model from Tencent Hunyuan built to do one job: decide what an AI system should remember. It’s a mixture-of-experts model with roughly 30 billion total parameters but only about 3 billion active at inference time, which is why “A3B” shows up in the name. Instead of generating chat responses, it slots into the memory-writing stage of frameworks like Mem0, MemOS, and Tencent’s own Hy-Memory, replacing the general-purpose LLM that normally handles memory extraction and updates. In Tencent’s published benchmarks, it outperforms both Claude Opus 5 and DeepSeek V4.1 Flash on the averaged memory-writing tasks across four different harnesses.
TL;DR
- Hy-Memory-30B-A3B is a memory-writer model, not a chatbot, designed to replace the general LLM that handles the write path in memory frameworks like Mem0 and MemOS.
- It uses a mixture-of-experts design with around 30B total parameters and roughly 3B active per forward pass, keeping inference costs closer to a small model despite the larger total weight count.
- Tencent trained it in two stages, starting with on-policy distillation from a teacher model and following with reinforcement learning that scores entire memory banks rather than single outputs.
- It beat Claude Opus 5 and DeepSeek V4.1 Flash on the overall cross-harness average (55.49 vs 54.62 and 53.64 respectively) across Mem0, Hy-Memory, MemOS, and the held-out EverMemOS framework.
- The model generalizes to frameworks it never trained on: EverMemOS was excluded from training data entirely and Hy-Memory-30B-A3B still posted the top score there.
- It’s open weight under Apache 2.0 and ships with vLLM deployment instructions, an OpenAI-compatible API, and a “no-think” mode for faster, non-reasoning extraction calls.
- Weaker performance showed up on retail-style tasks (the “Shop.” column), where several larger models, including Qwen variants, edged it out on specific subsets.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Why does AI need a dedicated memory-writer model?
Most AI agents that claim “long-term memory” work the same way: a conversation happens, a general-purpose LLM reads it, extracts facts worth keeping, and writes them into a persistent store (a vector database, a graph, or some hybrid). Frameworks like Mem0 and MemOS are built to manage that store, but they still rely on whatever LLM you plug in to actually decide what’s worth remembering and how to phrase it.
That’s a mismatch. A model tuned to write engaging prose or solve coding problems isn’t necessarily tuned to make good judgment calls about which details matter six conversations later, how to merge a new fact with a contradicting old one, or how to keep a memory bank from bloating with redundant entries. Tencent’s pitch with Hy-Memory-30B-A3B is that this write path deserves its own specialized model, the same way retrieval and reranking already have their own specialized models in most production RAG stacks.
How was it trained?
Tencent used a two-stage process described in the model’s technical report (arXiv:2610.22196).
Stage one: Memory On-Policy Distillation (OPD). The model is trained inside the same harness configurations it will later be deployed in, learning from dense supervision on “student-written state chains,” meaning a larger teacher model grades and corrects the memory states the smaller model itself generates, rather than just handing over static examples.
Stage two: Attributed Memory Reinforcement Learning. Here, the training process looks at entire memory banks built across a rollout, not just individual outputs. Each chunk of a memory-writing session gets a score, credit is assigned at the chunk level, and the highest-scoring resulting bank gets carried forward as the parent for the next window. This is a meaningfully different setup from typical RLHF, which usually scores a single response. Scoring a whole evolving memory bank pushes the model to optimize for how memories compound and interact over a long session, not just whether one extraction step looked reasonable in isolation.
How does it perform against Claude Opus 5 and DeepSeek V4.1 Flash?
Tencent evaluated all memory writers under identical conditions: same harness version, same prompts, same retrieval setup, same answer model and judge. Only the memory-writing model changed between runs. Each writer started from an empty memory state and built its own bank from scratch, and scores came from a fixed 20% subsample across six benchmarks: LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench.
On the overall cross-harness average, the ranking was:
- Hy-Memory-30B-A3B: 55.49
- Claude Opus 5: 54.62
- DeepSeek V4.1 Flash: 53.64
- Hy3 (Tencent’s general model): 51.63
- Qwen3.6-35B-A3B: 50.37
Breaking it down by harness, Hy-Memory-30B-A3B led on Mem0 (54.02), Hy-Memory (60.55), and MemOS (50.51), and also topped the held-out EverMemOS harness (56.89), which it never saw during training. That last result is the more interesting one: generalizing to an unseen framework suggests the model learned something about memory-writing behavior broadly, rather than overfitting to the specific prompts and formats of the frameworks it trained against.
One coffee. One working app.
You bring the idea. Remy manages the project.
The picture isn’t uniformly dominant, though. Look at the “Shop.” (Shopping PS) column across every table and Hy-Memory-30B-A3B rarely tops it. Qwen3.6-35B-A3B hit 20.00 on Mem0’s shopping metric versus Hy-Memory’s 18.33, and Claude Opus 5 led shopping scores in both the Hy-Memory and EverMemOS harnesses. Shopping-related memory, which likely involves tracking preferences, sizes, and purchase history across noisy conversational signals, seems to be a harder case where general-purpose models still hold an edge in some configurations.
Is Hy-Memory-30B-A3B worth deploying over a general LLM?
For teams already running Mem0, MemOS, or similar memory frameworks, the case comes down to three things: cost, latency, and whether the benchmark gains translate to your actual data.
On cost and latency, the active-parameter count (~3B) matters more than the total (~30B) for inference speed and serving economics, since mixture-of-experts models only route tokens through a subset of experts per forward pass. That makes Hy-Memory-30B-A3B considerably cheaper to run than Claude Opus 5 or even DeepSeek V4.1 Flash for high-volume memory-writing calls, which often run on every single conversation turn in a production agent.
On accuracy, the margins in Table 1 are real but not huge. A 55.49 vs 54.62 overall average is a win, but it’s close enough that teams with workloads heavy in retail-style or preference-tracking memory should check the Shopping PS numbers specifically before assuming a blanket improvement. The model also ships open weight under Apache 2.0, which matters for teams that want to self-host the memory layer rather than route sensitive conversation data through a third-party API purely for the extraction step.
How do you actually run it?
The model card documents deployment via vLLM (tested with version 0.31.0), served with an OpenAI-compatible API. A typical call sets up a system prompt instructing the model to extract explicit facts and preferences, return them as a JSON list of self-contained statements, and avoid inferring anything not stated outright. The model also supports a “no-think” mode through a reasoning_effort field in the chat template, useful for keeping latency down on what should be a fast, mechanical extraction task rather than a reasoning-heavy one.
Weights, config files, and the tokenizer are hosted on Hugging Face under tencent/Hy-Memory-30B-A3B, split across 51 safetensors shards, with the framework code and evaluation scripts available on GitHub under Tencent-Hunyuan/Hy-Memory.
Frequently Asked Questions
What does “A3B” mean in the model name?
It refers to the mixture-of-experts architecture: the model has around 30 billion total parameters but activates only about 3 billion of them for any given input, which keeps inference cost and latency closer to a much smaller dense model.
Which memory frameworks does Hy-Memory-30B-A3B work with?
It’s designed to replace the general LLM inside Mem0, MemOS, and Tencent’s own Hy-Memory framework (covering chat, working memory, and coding scenarios). It also transferred successfully to EverMemOS, a framework it wasn’t trained on.
How was it benchmarked against Claude Opus 5 and DeepSeek V4.1 Flash?
Tencent ran all three models as drop-in memory writers inside identical harness setups, fixing prompts, retrieval, and the downstream answer/judge models, then scored each on a fixed 20% subsample across six memory benchmarks including LoCoMo and LongMemEval.
Is Hy-Memory-30B-A3B better in every scenario?
No. It led on overall averages and three of four harnesses, plus the held-out EverMemOS test, but some retail/shopping-preference metrics favored Claude Opus 5 or Qwen3.6-35B-A3B instead.
Is the model open source?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Yes. It’s released under the Apache 2.0 license on Hugging Face, with accompanying framework code and evaluation utilities on GitHub and a technical report on arXiv (2610.22196).
