Tencent Hy-Memory-30B-A3B: An Open 3B-Active Model That Writes Agent Memory

Tencent Hunyuan has put a very specific kind of model on Hugging Face: Hy-Memory-30B-A3B, an Apache-2.0 mixture-of-experts model with roughly 3B active parameters whose only job is to decide what an AI agent should remember. According to the model card, it drops into Mem0, MemOS and Tencent's own Hy-Memory framework in place of a general-purpose LLM, and on Tencent's benchmark matrix it averages 55.49 versus 54.62 for Claude Opus 5. Those are Tencent's numbers, not an independent test, and the post below separates what the release shows from what it does not.

Quick reference

What is a "memory writer" and why does it need its own model?

Most agent memory systems have two halves. The read path retrieves stored facts when a new request arrives. The write path decides, after every conversation chunk or work session, what is worth keeping, how precisely to phrase it, and whether it contradicts or updates something already stored.

The write path is where memory quality is won or lost. If the writer stores too much, retrieval fills with noise. If it stores too little or paraphrases loosely, the agent forgets a dietary restriction or misremembers a codebase convention. Frameworks such as Mem0 and MemOS usually call a general-purpose frontier model for this step, which is expensive when it runs after every turn.

Hy-Memory-30B-A3B targets that cost. Tencent describes it as a compact long-term-memory writer: with about 3B parameters active per token, serving it is closer to running a small model than a frontier one, while it is meant to match frontier quality on this one task. That framing fits a wider pattern, where a small specialised model competes with larger general ones on one measured task, and it sits alongside the argument in why agents may need documentation more than memory.

Tencent Hy-Memory-30B-A3B as an agent long-term memory store, shown as spools of thread kept in a small drawer

How was it trained?

The model card describes a two-stage recipe, both stages run inside the same deployed harness configurations that the model will later serve:

1. Memory On-Policy Distillation (OPD). A stronger teacher supplies dense supervision on state chains that the student itself wrote. "On-policy" means the student's own memory banks, not a teacher's idealised ones, are what get corrected, which keeps training close to deployment conditions.
2. Attributed Memory Reinforcement Learning. The resulting memory banks are examined, and credit is assigned at the chunk level. In the model card's figure, the best rollout can differ from chunk to chunk, and the bank with the highest summed chunk score becomes the parent for the next window.

The chunk-level credit matters. Memory writing is a long-horizon task: a bad decision in chunk 3 may only show up as a wrong answer after chunk 40. Scoring each chunk's contribution is Tencent's answer to that credit-assignment problem. Whether it generalises is exactly what the held-out EverMemOS test is meant to probe.

What do the benchmark numbers say?

Tencent evaluates by changing only the memory writer. The harness version, prompts, retrieval configuration, answer model and judge are fixed. Each harness score averages six metrics equally: LoCoMo Judge, LongMemEval, PersonaMem-v2, Travel SPS, Shopping PS and Code Memory Bench. The overall score averages four harnesses equally. According to the report, every DeepSeek Flash reference denotes DeepSeek V4.1 Flash, and every benchmark runs on a fixed 20 percent subsample for efficiency.

Source: Tencent Hunyuan model card, Table 1. Six-metric means on fixed 20 percent subsamples.

Three things stand out.

The small model gains most over its own base. Hy3-A3B Base scores 47.66 and the tuned model scores 55.49, a gain of about eight points. That is the clearest evidence in the table that the training recipe, not the architecture alone, did the work. The older Qwen3-30B-A3B sits at 44.37.

The lead over Opus 5 is thin. On the Hy-Memory harness the two are 60.55 versus 60.53, and on MemOS 50.51 versus 50.50. Those are ties. The real separation is on EverMemOS (56.89 against 54.04) and Mem0 (54.02 against 53.43). With subsampled benchmarks, differences of this size could move with a different sample.

Frontier models still win individual benchmarks. In the Mem0 detail table, Opus 5 scores 69.47 on LoCoMo against 59.54 for Hy-Memory, and 70.74 against 67.13 on Code Memory Bench. Hy-Memory leads on LongMemEval (73.00 versus 66.00) and Travel (54.51 versus 42.98). So the average hides a different profile: stronger at some memory tasks, weaker at others.

Note also that the Shopping PS scores are low for every model (roughly 14 to 24), which suggests that benchmark is either hard or noisy and contributes little to separating writers.

Why EverMemOS is the number to watch

Three of the four harnesses were part of training. EverMemOS was not. On it, Hy-Memory-30B-A3B scores 56.89, ahead of DeepSeek V4.1 Flash at 55.47 and Opus 5 at 54.04. If the result holds under independent testing, it supports the claim that the writer learned something general about memory rather than overfitting to three frameworks' prompt formats.

The caution is that the benchmark composition is Tencent's choice. Another team using different tasks, or full datasets instead of a 20 percent slice, might see a different ranking. For a comparable story about self-reported agent-memory scores, see our look at Garry Tan's GBrain evals and the self-graded benchmark problem.

How to try it

The model card gives a short quickstart. You serve the weights with vLLM (Tencent tested v0.31.0) under the name hy-memory, then call the OpenAI-compatible API. The prompt in the example asks the model to extract explicit user facts and preferences and return a JSON object with a list of memory strings, with temperature 0 and the chat template argument reasoning_effort set to no_think.

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="hy-memory",
    messages=[
        {"role": "system", "content": "Extract explicit user facts and preferences. Return only a JSON object with a memories list of strings."},
        {"role": "user", "content": "I live in Seattle, mainly use Python, and I am vegetarian."},
    ],
    temperature=0.0,
    max_tokens=512,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}},
)
print(response.choices[0].message.content)

The chat template also supports high and low reasoning effort, and no_think is the default when none is given. For production use, the evaluation utilities and framework code live in the Hy-Memory GitHub repository. We have not run the model ourselves for this post, so the quickstart above reflects the model card, not our own testing.

What this means if you build agents

• Cost. If your agent calls a frontier model after each turn only to update memory, a 3B-active writer could cut that bill. The saving depends on your hardware: a 30B-total model still needs the memory to hold all experts, even though only a fraction compute per token.
• Privacy and control. Memory writers see the most sensitive text an agent handles: preferences, names, work history. An Apache-2.0 model you host yourself keeps that text off third-party APIs, a point that matters for the enterprise deployments discussed in our Cognition Devin memory and open spec coverage.
• Framework lock-in. Mem0, MemOS and EverMemOS being supported means you can swap the writer without rebuilding the retrieval side.
• Evaluation hygiene. If you adopt it, run your own conversations through both the frontier writer and this one and compare the stored banks by hand. Memory quality is domain-specific, and a public average will not tell you whether it keeps your users' constraints.

Limits and open questions

• Self-reported. All figures come from Tencent's report. No external replication was visible at publication.
• Subsampled evaluation. Only 20 percent of each benchmark was used, which widens the error bars on sub-point differences.
• Judge and answer model fixed. Good for fairness, but results may shift with other answer models. The answer model and judge are not the writer, and the report holds them constant.
• Writer only. The model does not do retrieval. A great writer paired with a weak retriever still gives poor recall.
• Fresh release. The weights had no downloads yet when we checked and a handful of likes, so community feedback on edge cases, languages beyond English and Chinese, and failure modes does not exist yet.
• Operational risk. Any model that rewrites stored memory can also corrupt it: a wrong update silently overwrites a right fact. Keep history and audit logs. If your agents act on memory with real permissions, a runtime guard such as AgentBeam can stop dangerous actions that a bad memory might trigger.

Where it fits in the Tencent and agent landscape

Tencent has been shipping a steady stream of focused open models, such as the Youtu-Parsing Omni 5B document and audio parser, rather than only chasing a single flagship. Hy-Memory follows the same logic: pick a narrow job inside the agent stack, train a small model on it with in-harness reinforcement learning, publish weights and an evaluation harness. For readers building computer-use or coding agents, the memory layer is increasingly the differentiator, as the OpenAI contracting agents with computer use story shows when an agent must carry context across many sessions.

The practical takeaway: treat Hy-Memory-30B-A3B as a credible, cheap candidate to test as your memory writer, not as proof that a small model has overtaken frontier models on memory in general.

Related reading

• Agents don't need memory, they need documentation
• Cognition Devin dreaming memory and the open memory spec
• Garry Tan's GBrain evals: a self-graded agent-memory benchmark
• Tencent Youtu-Parsing Omni 5B
• OpenAI contracting agents and computer use
• Claude Opus 5.5 launch, benchmarks and pricing

Specifications and scores are as published by Tencent Hunyuan on October 16, 2026 and may change as the repository and report are updated.