Tencent’s Hy-Memory is built to write the memories AI agents keep

Tencent’s Hy-Memory-30B-A3B is a specialist model for deciding what an AI agent should remember, rather than a chatbot for answering questions. OrcaRouter says the model’s weights and technical report appeared on 15 October, offering developers a memory-writing model with 30 billion total parameters and roughly 3 billion active per token.

Tencent AI Watch analysis

What happened

The model is designed to add, update and organise information in persistent memory banks used by agent frameworks. OrcaRouter says it was trained across Mem0, MemOS and Tencent’s Hy-Memory, then evaluated on EverMemOS, which was held out of training. The report says the model is released under Apache 2.0 and can be served through vLLM behind an OpenAI-compatible endpoint.

There is a substantial hardware catch: OrcaRouter says the available BF16 checkpoint is about 60GB, with no quantised version identified in its account. The report also describes a 262,144-token context window. Its central performance claim is a 55.49 average across four memory benchmarks, just ahead of the cited Claude Opus 5 score of 54.62. Those results come from Tencent’s evaluation, which used a 20 per cent subsample, and have not been independently reproduced, according to OrcaRouter.

Why it matters

Agent systems often use a general-purpose model to decide what belongs in long-term memory. A model trained specifically for that job could make the memory-writing step more consistent and, if it proves cheaper in real use, less expensive to run repeatedly. Testing on a framework left out of training is a useful part of the evaluation design, though it is not a substitute for another team reproducing the result.

For now, this looks more like infrastructure for teams with serious hardware than a model most developers can casually try. The reported checkpoint size and lack of quantisation put practical access, not just benchmark rank, squarely in the conversation.

Our read

The interesting idea is not another model that can remember a conversation. It is a model whose job is to write the memory bank in the first place. That is a credible specialism, and Tencent’s reported cross-framework test is worth attention. But the headline benchmark remains a claim to test, not a settled win; the first useful evidence would be independent results and a lighter checkpoint.

What to watch

• Whether independent teams reproduce the held-out EverMemOS result.
• Whether Tencent or the community releases quantised weights.
• Whether real deployments publish memory-writing costs and error rates.

Discussion spark: Would you trust a specialist model to decide what an AI agent remembers, or should that job stay with the model handling the conversation?

not affiliated with or endorsed by Tencent AI