Tencent Hunyuan has released Hy Memory 2.0, an open-source memory framework for AI agents. The release includes a specialized model, the memory framework itself, and the MemHub application, all of which can be used independently or as a unified solution.
The framework introduces specific capabilities for chat, work, and code memory. In benchmark tests, Hy Memory 2.0 outperformed several open-source alternatives, including Mem0 and EverOS. Specifically, it achieved a 64.2% accuracy rate in code Q&A and a 72.6% accuracy rate in chat memory tasks.
The specialized Hy-Memory-30B-A3B model, which features approximately 3B active parameters, was trained on three frameworks. It achieved an average score of 55.49% across four evaluation benchmarks, surpassing Claude Opus 5, which scored 54.62%.
MemHub, a local memory management application, is available via CLI and a Mac app. It allows users to import session history from various tools and sync new data in real-time. It supports 10 AI tools, including Cursor, WorkBuddy, and DeepSeek Harness.
The framework is currently utilized within Tencent for various internal business sectors, including finance, knowledge management, and gaming. The SDK and MemHub are available on GitHub, while model weights are hosted on Hugging Face.
