Tencent's Hy-Memory-30B-A3B: A Model Built Just to Write Memory
Tencent's Hy-Memory-30B-A3B is a dedicated memory-writer model that outperforms general LLMs inside Mem0, MemOS and Hy-Memory frameworks.
What is Hy-Memory-30B-A3B?
Hy-Memory-30B-A3B is a model from Tencent Hunyuan built for one job: deciding what a long-term memory system should store, how precisely it should store it, and how to update that stored memory as new conversation comes in. It has roughly 3 billion active parameters and is meant to replace the general-purpose LLM that memory frameworks like Mem0, MemOS, and Tencent’s own Hy-Memory currently use to do that job. It’s not a chat model or a reasoning model in the usual sense. It’s a specialist for the “write path” of memory systems, trained through a two-stage process Tencent calls Memory On-Policy Distillation followed by Attributed Memory Reinforcement Learning.
TL;DR
- Hy-Memory-30B-A3B is a memory-writer model, not a general chatbot, designed to slot into existing memory frameworks and handle what to remember and how to update it.
- It directly replaces the general LLM inside Mem0, MemOS, and Hy-Memory, which today typically call a general-purpose model to do this same job less efficiently.
- On a four-harness benchmark matrix (Mem0, Hy-Memory, MemOS, and the held-out EverMemOS), it posted an average score of 55.49, edging out Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64).
- The model generalizes to a memory framework it never trained on (EverMemOS), scoring 56.89 there, which suggests the skill it learned is about memory writing in general rather than one harness’s quirks.
- Training happens in two stages: distillation from a stronger teacher on memory-writing traces, then reinforcement learning that scores entire memory banks and assigns credit to the chunks that built them.
- It’s deployable today via vLLM with an OpenAI-compatible API, and Tencent published the weights, framework code, and a technical report alongside it.
- Despite its relatively small active parameter count, it beat much larger general models like Qwen3-30B-A3B and Hy3 across every harness tested.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Why does memory writing need its own model?
Most “AI with memory” products today work by bolting a memory layer onto a general LLM. Frameworks like Mem0 and MemOS extract facts from a conversation, decide whether to add, update, or discard entries in a memory bank, and later retrieve relevant memories to feed back into a prompt. The extraction and update step, the “write,” is usually handled by calling a general-purpose model like GPT-4, Claude, or DeepSeek with a prompt that says something like “pull out the facts worth remembering.”
That works, but it’s an awkward fit. A general model wasn’t trained to make these specific decisions well: when two facts conflict, which one wins? When a new message slightly updates an old preference, do you overwrite or append? How much detail is worth keeping versus how much is noise? These are narrow, repeatable decisions, and training a model specifically on them, rather than hoping a general chat model happens to be good at it, is the premise behind Hy-Memory-30B-A3B.
Tencent’s benchmarks back this up directly. A smaller model trained for exactly this task beat multiple general-purpose frontier-class models on the metrics that matter for memory quality, across several different frameworks.
How was Hy-Memory-30B-A3B trained?
The model goes through two training stages, both run inside the actual deployed memory harnesses rather than on some abstracted synthetic task.
The first stage is Memory On-Policy Distillation. A stronger teacher model supervises the student (Hy-Memory-30B-A3B) on the memory “state chains” the student itself writes, meaning the student generates its own memory-writing attempts and gets dense corrective feedback on them, rather than just imitating teacher outputs blindly.
The second stage is Attributed Memory Reinforcement Learning. Here the system looks at entire memory banks built across a rollout, scores them, and assigns credit at the level of individual chunks, not just the final output. Because conversations and memory updates happen over long windows, the best “memory bank” from one window becomes the starting point (the parent) for the next window. The reinforcement signal is normalized within each chunk-level group, so the model can learn that one written chunk was a good contribution even if the overall rollout wasn’t the best one, or vice versa.
The effect of this design is a model trained specifically to track evolving memory state over long multi-turn interactions, which is a different skill than being generally good at conversation or reasoning.
How was it benchmarked, and how good is it really?
Tencent evaluated Hy-Memory-30B-A3B by swapping it in as the memory writer inside four harnesses: Mem0, Hy-Memory, MemOS, and EverMemOS (a framework the model never saw during training). Everything else in each setup, the harness version, the prompts, the retrieval configuration, the answer model, and the judge, was held fixed. Only the memory-writer model changed between runs, which makes the comparison a reasonably clean test of writer quality in isolation.
Each harness score is an equal-weighted average of six metrics: LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS, and Code Memory Bench. Evaluation ran on a fixed 20% subsample of each benchmark for computational efficiency.
The headline numbers: Hy-Memory-30B-A3B scored an overall average of 55.49 across the four harnesses, ahead of Claude Opus 5 (54.62), DeepSeek V4.1 Flash (53.64), Tencent’s own Hy3 (51.63), and Qwen3.6-35B-A3B (50.37). It also clearly outperformed smaller or base comparisons like Hy3-A3B Base (47.66) and Qwen3-30B-A3B (44.37).
Per-harness, it topped or tied for the top score on Hy-Memory (60.55) and the held-out EverMemOS (56.89), and came very close to the best score on Mem0 and MemOS. The EverMemOS result matters most for anyone wondering if this model only works well on frameworks it was explicitly tuned for. Since EverMemOS was held out from training entirely, a strong score there is evidence the model learned general memory-writing behavior rather than memorizing quirks of specific harnesses.
Looking at individual sub-metrics, performance isn’t uniformly dominant. Claude Opus 5 still leads on LoCoMo in the Mem0 harness (69.47 versus 59.54), and DeepSeek V4.1 Flash edges it out on some Code Memory Bench scores. But averaged across the full matrix of six metrics and four harnesses, Hy-Memory-30B-A3B comes out on top.
How do you actually deploy it?
Tencent ships this as an open-weight model with a working deployment path. It’s designed to run with vLLM (tested against version 0.31.0), served with an OpenAI-compatible API. A typical integration involves sending the conversation to the model with a system prompt instructing it to extract explicit facts and preferences, return them as concise self-contained statements, and avoid inferring information that wasn’t actually stated. The model supports a “no-think” mode for faster, more direct extraction without a visible reasoning trace, which is useful for a task like memory writing where you generally want speed and a clean JSON output rather than a long chain of thought.
Tencent released the framework code and evaluation utilities on GitHub alongside the Hugging Face model weights and a technical report, so teams building or already using Mem0, MemOS, or similar memory layers can test a drop-in replacement for their current writer model.
Is a dedicated memory-writer model worth switching to?
For teams running production memory systems with Mem0, MemOS, or Hy-Memory, the benchmark results suggest a real quality gain is possible by swapping the generic LLM call in the write path for a specialized one, without necessarily increasing cost given the model’s relatively small active parameter footprint. The generalization result on EverMemOS is the more interesting signal for teams on frameworks Tencent didn’t explicitly test against: it implies the skill transfers rather than being tightly coupled to one harness’s prompt format. Whether it’s worth switching depends on how much memory quality currently bottlenecks your application versus retrieval quality or context window limits, which this model doesn’t address.
Frequently Asked Questions
What does “A3B” mean in the model name?
It refers to the model’s active parameter count during inference, roughly 3 billion, consistent with the pattern Tencent uses for its Hunyuan (Hy) model family, which likely uses a mixture-of-experts architecture where only a subset of parameters activates per token.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Which memory frameworks does Hy-Memory-30B-A3B work with?
It’s designed to replace the general LLM used by Mem0, MemOS, and Tencent’s own Hy-Memory framework (covering chat, working, and coding memory scenarios). It also generalized well to EverMemOS, a framework it wasn’t trained on.
How is it different from a regular chat model being used for memory tasks?
Most memory frameworks today call a general-purpose chat model to extract and update memories, using prompting alone to steer that behavior. Hy-Memory-30B-A3B is trained specifically for this task through distillation and reinforcement learning focused on memory bank quality, rather than general conversational ability.
Is Hy-Memory-30B-A3B open source?
Yes. Tencent published the model weights on Hugging Face, the framework and evaluation code on GitHub, and a technical report detailing the training methodology.
What hardware or software do you need to run it?
Tencent’s documentation describes deployment through vLLM (tested with version 0.31.0), exposed via an OpenAI-compatible API, which makes it straightforward to call from existing applications that already talk to OpenAI-style endpoints.
