Tencent Releases Hy-Memory Memory-Writing Model, Trained with Attribution-Based Reinforcement Learning to Retain Information Across Frameworks

Tencent Hunyuan announced Hy-Memory 2.0 on October 16, releasing open-source resources across three layers: models, memory frameworks, and applications. Hy-Memory-30B-A3B is designed specifically for the write process in external memory systems: it selects information worth retaining from interaction histories and turns it into memory that can be updated over time. The model weights are released under the Apache 2.0 license. The model has about 30 billion total parameters, with about 3 billion active parameters. Official announcement, model card

The technical focus is turning “how to write memory” into a trainable capability. The research team first uses Memory On-Policy Distillation to teach the model the write protocols of different frameworks using memory states the model generated itself. It then applies attribution-based memory reinforcement learning to optimize whether retained information supports later answers, using step-by-step probing, cross-block reward attribution, and controls against over-storage. Training also carries the memory bank written by the model into the next window, making state evolution more similar to deployment. Technical paper

The official evaluation holds retrieval settings, the answer model, and the judge constant, changing only the memory-writing model; each model starts with an empty memory bank. Tests cover Mem0, MemOS, Hy-Memory, and EverMemOS, which was not included in training. Averaged equally across four frameworks and six metrics, Hy-Memory scores 55.49, compared with 54.62 for Claude Opus 5. However, the main table uses only a fixed 20% test subset, and the average lead does not mean better performance on every subtask. For example, Opus 5 scores higher on Mem0’s LoCoMo and code-memory tests. Evaluation method and per-task results

For agent-system developers, this provides an experimental entry point for independently swapping in a memory-writing model and observing how information-retention strategies affect long-term tasks. The model card includes deployment instructions tested with vLLM 0.31.0 and an example using the OpenAI-compatible API. Next, developers should validate the full test set, long Traditional Chinese conversations, and information-correction scenarios, while also measuring write latency, memory-bank size, and service costs. The publicly reported scores are not yet enough to show that real-world deployment uses fewer resources. Deployment instructions