Hy-Memory-30B-A3B vs Claude Opus 5 and DeepSeek on Memory Tasks
Tencent's small memory-writer model Hy-Memory-30B-A3B edges out Claude Opus 5 and DeepSeek V4.1 Flash on memory benchmarks. Here's what the numbers show.
What is Hy-Memory-30B-A3B and why does it beat bigger models at memory tasks?
Hy-Memory-30B-A3B is a specialized “memory writer” model from Tencent Hunyuan, built with roughly 3 billion active parameters, that decides what a long-term memory system should store, how it should be phrased, and when old entries need updating. In Tencent’s published benchmarks across four memory harnesses (Mem0, Hy-Memory, MemOS, and the held-out EverMemOS), it scored an average of 55.49, narrowly ahead of Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64. The takeaway: a small model trained specifically for the memory-writing job can match or beat general-purpose frontier models at that one job.
TL;DR
- Hy-Memory-30B-A3B is not a chat model. It’s a drop-in replacement for the LLM that sits inside memory frameworks like Mem0, MemOS, and Tencent’s own Hy-Memory, handling only the “write path” that decides what gets saved to long-term memory.
- It beat Claude Opus 5 and DeepSeek V4.1 Flash on average across four memory harnesses, scoring 55.49 versus 54.62 and 53.64 respectively, despite having a small fraction of the active parameters those models presumably carry.
- The training recipe has two stages: Memory On-Policy Distillation (OPD), where a stronger teacher supervises the model’s own memory-writing attempts, followed by Attributed Memory Reinforcement Learning, which scores and credits individual memory chunks rather than whole conversations.
- It generalizes to frameworks it never saw during training. EverMemOS was held out entirely, and Hy-Memory-30B-A3B still posted the top score there (56.89), suggesting the skill transfers rather than overfitting to one harness’s quirks.
- Results are uneven across sub-benchmarks. Claude Opus 5 still wins outright on raw LoCoMo judge scores in most harnesses, and DeepSeek V4.1 Flash leads on some Code Memory Bench and Travel SPS numbers, so “best overall” doesn’t mean “best at everything.”
- The comparison isolates one variable. Tencent’s evaluation setup fixes the harness, prompts, retrieval config, answer model, and judge, and only swaps out the memory writer, which makes the head-to-head reasonably clean as benchmark comparisons go.
- It’s open and deployable now, with weights on Hugging Face and a vLLM-based quickstart that serves it through an OpenAI-compatible API.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
How does a “memory writer” model differ from a regular LLM?
Most benchmark conversations about LLMs are about generation: can the model answer a question, write code, or reason through a math problem. Memory writer models do a narrower and less visible job. In systems like Mem0 or MemOS, a conversation happens between a user and an assistant, and a separate background process decides what from that conversation is worth remembering long-term, how to phrase it so it’s useful later, and whether a new fact should overwrite or merge with something already stored. That decision-making is the “write path,” and it’s normally handled by whatever general-purpose LLM the framework happens to call.
Hy-Memory-30B-A3B is built to replace that general-purpose call with a model trained specifically for the job. It doesn’t answer the user’s question. It watches the conversation and manages the memory bank. Tencent designed it to slot into Mem0, MemOS, and its own Hy-Memory framework (which covers chat, working memory, and coding scenarios) without needing those frameworks to be rebuilt around it.
How was it trained?
The model goes through two stages, both run inside the same deployed harness configurations it will later be evaluated in, rather than on a generic pretraining corpus disconnected from the actual task.
The first stage is Memory On-Policy Distillation (OPD). The model writes its own memory “state chains” (the sequence of memory banks it produces as a conversation unfolds), and a stronger teacher model supplies dense supervision on those self-generated chains. This is “on-policy” because the student learns from correcting its own outputs rather than just imitating teacher-generated examples from scratch.
The second stage is Attributed Memory Reinforcement Learning. Here, the system examines the memory banks produced across multiple rollouts, assigns credit at the level of individual chunks rather than whole conversations, and carries the best-performing bank forward into the next window. Advantages get normalized within each chunk-level group, meaning the “best” rollout isn’t fixed. It can change from chunk to chunk depending on which version of the memory bank scored best at that point. The final bank selected is the one with the highest cumulative chunk score, and that becomes the starting point for the next stretch of conversation.
How do the benchmark numbers actually break down?
Tencent evaluated seven memory writers (Hy-Memory-30B-A3B, Claude Opus 5, DeepSeek V4.1 Flash, Hy3, Qwen3.6-35B-A3B, Hy3-A3B Base, and Qwen3-30B-A3B) across four harnesses: Mem0, Hy-Memory, MemOS, and EverMemOS. Each harness score averages six metrics: LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench. All of this runs on a fixed 20% subsample of each benchmark for computational efficiency, and only the memory writer changes between runs, everything else (prompts, retrieval, the answer model, the judge) stays fixed.
The headline numbers:
| Memory writer | Mem0 | Hy-Memory | MemOS | EverMemOS | Average |
|---|---|---|---|---|---|
| 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 |
The margins are thin. Hy-Memory-30B-A3B and Claude Opus 5 are essentially tied on the Hy-Memory and MemOS harnesses (differences under half a point). The bigger separation shows up on Mem0 and EverMemOS, where Hy-Memory-30B-A3B pulls ahead by roughly 0.6 and 2.8 points respectively.
Looking at the sub-metrics tells a more mixed story. Claude Opus 5 wins outright on raw LoCoMo judge scores in the Mem0 and MemOS harnesses (69.47 and 63.74 versus Hy-Memory-30B-A3B’s 59.54 and 62.98). DeepSeek V4.1 Flash takes the top Code Memory Bench score in Mem0 and MemOS, and the top Travel SPS score on EverMemOS. Hy-Memory-30B-A3B’s strength isn’t dominating every category, it’s being consistently competitive across all of them, which is what drives the higher averages.
Is Hy-Memory-30B-A3B actually worth using over a general-purpose model?
For teams building memory-augmented agents, the practical argument isn’t just “it scores higher.” It’s that a ~3B active-parameter model matching or beating frontier-scale models on this specific task implies a meaningful efficiency trade. Running a large general-purpose model purely to decide what to write into a memory bank is expensive and often overkill for a task that’s narrower than open-ended generation. If a much smaller specialized model performs comparably, that’s a real cost and latency win for any system doing memory writes continuously in the background.
The caveat is that these results come from Tencent’s own benchmark suite and evaluation harnesses, run on 20% subsamples, with methodology published in their technical report rather than independently replicated across the board. The gaps between top models are also small enough that framework choice, prompt design, and retrieval configuration likely matter as much as which model does the writing. Anyone evaluating this for production use should expect to need their own testing against their actual memory framework and data, not just the published averages.
The model also doesn’t generalize to being a general-purpose assistant. It’s trained and shaped around the narrow task of extracting, compressing, and updating memory content, demonstrated in its quickstart example where it extracts structured facts (name, location, job, dietary preference) from a short conversation into a clean JSON list. That’s a different skill set than open-ended chat or coding.
Frequently Asked Questions
What does “A3B” mean in Hy-Memory-30B-A3B?
It refers to the model’s architecture, indicating around 3 billion active parameters within a larger total parameter count, consistent with a mixture-of-experts design where only a subset of parameters activate per inference call.
Which memory frameworks does Hy-Memory-30B-A3B work with?
It’s designed to plug into Mem0, MemOS, and Tencent’s own Hy-Memory framework (covering chat, working, and coding scenarios), and it also transferred successfully to EverMemOS, a framework it was never trained on.
Did Hy-Memory-30B-A3B beat Claude Opus 5 on every benchmark?
No. It won on the overall average and on three of the four harness averages, but Claude Opus 5 posted higher raw LoCoMo judge scores in several harnesses. The comparison is close, not a clean sweep.
Is Hy-Memory-30B-A3B open source?
Yes. Model weights and configuration are published on Hugging Face, the framework and evaluation code are on GitHub, and a technical report is available on arXiv.
How do you deploy Hy-Memory-30B-A3B?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Tencent’s quickstart uses vLLM to serve the model with an OpenAI-compatible API, including a “no-think” reasoning mode for fast, direct extraction of facts and preferences from conversation text.
