Tencent open-sources a 30B model that only writes agent memory, under Apache 2.0 — its 0.87-point lead over Opus 5 is an in-house figure from 20% of the eval data

Tencent Hunyuan has released Hy-Memory-30B-A3B on Hugging Face. The model is built only to handle long-term memory for AI agents. It is licensed under Apache 2.0 and uses a mixture-of-experts (MoE) design: of its 30B total parameters, only 3B are active for each token. The model does not answer questions. Instead, it reads conversations and tasks, decides what to write into memory, and decides how to revise memories that already exist. That single job is all it was built for. Tencent says it outperforms Claude Opus 5 at this task. The evidence, however, is an in-house measurement on part of the evaluation data, and the margin is less than one point.

A model that handles only memory "writes"

Agent memory systems such as Mem0 and Letta usually hand a general-purpose LLM the work of reading conversations and extracting, storing and updating memories. Tencent has split off this write step and given it to a dedicated model. The authors of the paper state the goal as follows.

In other words, Tencent's position is that memory quality depends less on the database or retrieval system and more on the model upstream that decides what to write. That model decides what to keep, how precisely to record it, and how new facts should overwrite old memories. These questions are separate from retrieval accuracy in RAG. If the write step gets them wrong, wrong memories will come back no matter how well they are retrieved later.

Swapping out the general-purpose LLM inside Mem0 and MemOS

Hy-Memory-30B-A3B is meant to be used as a component of existing memory frameworks. The model card names the frameworks it can be swapped into and says it carries over to frameworks it was not trained on:

If Tencent's claims hold, developers already using Mem0 or MemOS only need to change the write model in their memory settings. The GitHub repository also contains Hy-Memory, Tencent's own memory framework. With 3B active parameters, the model costs far less to run than calling a frontier model's API for every write, and it is small enough to run on local GPUs. Write costs add up quickly for always-on agents and other setups that rewrite memory constantly. Tencent's design puts a cheap, dedicated model at that point, and the Apache 2.0 license puts few limits on commercial use.

What the 0.87-point gap actually measures

Tencent says Hy-Memory-30B-A3B scored 55.49 on this write task, against 54.62 for Claude Opus 5, a gap of 0.87 points. The number invites the headline that a 30B open-weight model has beaten a closed frontier model at a specialized task.

Three caveats apply. First, the comparison used a 20% sample of the evaluation data, not the full set. Scores can shift depending on how that sample is drawn, so it is unclear whether the ranking would hold on the full set. Second, Tencent ran the test itself, and the results depend on the settings and prompts it used to have Opus 5 perform the writes. Third, as of this writing, no third-party evaluation has reproduced the scores. A gap of 0.87 points could easily flip under different conditions like these.

In the same week, DeepSeek also self-reported that V4.1-Flash-1017 beat Opus 5 on a coding-agent score, and that claim has not been independently evaluated either. Open-weight labs keep picking Opus 5 as their benchmark and announcing, with their own numbers, that they beat it. Note also that the comparison is with Opus 5, a generation behind, not with the current top model, Opus 5.5.

Moving writes to a separate model means more to verify

When a dedicated model handles memory writes, its decisions shape everything the agent does afterward. A paper from early October (arXiv 2610.04083) reported that misaligned goals written into memory carried over to the next session in 58% of runs. The model that decides what to keep sits upstream of the agent's behavior.

That is why the 0.87-point difference in average score is not what adopters of Hy-Memory-30B-A3B should check. They should check whether it keeps false information and injected instructions out of memory, how reliably it overwrites outdated facts, and whether the results hold up on their own data. Because the weights are public under Apache 2.0, anyone can run these checks. Handing memory writes to a small, dedicated model is a design worth trying for Mem0 and MemOS users. But until a third party re-measures on the full evaluation data, the claim that it beats Opus 5 should be read as Tencent's own.