Hy-Memory-30B-A3B: Write and update agent memories

What it does

Hy-Memory-30B-A3B writes and updates persistent memories from conversations. Instead of answering as a general chat model, it decides which information to retain and how to represent it inside a separate memory framework. Tencent says it can replace the general model used for the memory-writing step in Mem0, MemOS, and Hy-Memory, and describes chat, working, and coding scenarios. It has 30 billion total parameters and about 3 billion active parameters. The model card describes training focused on writing attributed memories and updating stored information when new evidence arrives.

How to use it

Deploy the checkpoint with the card’s vLLM command and send conversation text to its OpenAI-compatible API. The example prompt asks it to extract explicit facts and preferences and return them as a JSON list, which a connected memory system can store.

Common use cases

• Send a conversation that mentions a person’s name, city, job, and preferences; receive those explicit details as memory entries.
• Add a later conversation with changed information and let the memory writer update the stored record.
• Connect the model as the memory-writing component in Mem0, MemOS, or Hy-Memory, while the surrounding framework handles storage and retrieval.