Tencent releases Hy-Memory-30B-A3B for writing long-term memory
The Apache 2.0 text-generation model has 30B total parameters and about 3B active parameters. Tencent trained it in two stages, Memory On-Policy Distillation and Attributed Memory Reinforcement Learning, to select what to preserve and generate memory updates. Tencent says it transfers to EverMemOS, which was excluded from training.
Visit original source