Tencent Hunyuan said it has upgraded its agent memory framework Hy Memory to version 2.0 and open-sourced the dedicated model Hy-Memory-30A3B, the memory framework, and the MemHub application. The three layers can run independently or be combined into a full memory solution.
In Tencent’s description, the framework is responsible for storing, updating and evolving three types of memory: chat, work and code. The company said the framework ranked first among open-source memory frameworks across four benchmarks. For the dedicated model, Tencent reported a score of 55.49 in evaluations across several memory systems, including Hy Memory, Mem0 and MemOS, compared with 54.62 for Opus5.
Broader memory coverage in Hy Memory 2.0
Tencent said Hy Memory 2.0 moves chat memory beyond scattered facts and turns it into complete event context. It also adds work memory and code memory. According to the company, the system can judge when prior experience applies and provide more precise support for long-horizon tasks.
MemHub adds local management and tool sync
MemHub is positioned as a local memory management application. It supports CLI and Mac, and can connect with 10 AI tools including WorkBuddy, Cursor and Codex. Tencent said the app supports historical memory import and real-time synchronization across tools.
Deployment in Tencent products
Tencent said the capability has already been deployed in Tencent Self-Selected Stocks, iWiki and Tencent News. In Tencent Self-Selected Stocks, multi-hop recall rose from 50% to 90%, while the overall score improved by 23.8%. In iWiki, answer accuracy using memory alone was 48.6% higher on a relative basis than using all original dialogues. For Tencent News, the average score in long-term preference evaluation was 91.68.