Tencent Hy released Hy Memory 2.0 as open source, pitching it as "your private memory steward" for AI agents. The release has three pieces: a 30B-A3B model called Hy-Memory-30B-A3B, a framework that handles chat, work and code memories, and MemHub, a CLI and Mac app that syncs memory across more than 10 AI tools. Tencent says Hy Memory 2.0 is the #1 open-source system on 4 benchmarks, and that Hy-Memory-30B-A3B beats Opus 5 across major frameworks.
From Tencent Hy's announcement.
The three pieces
The model. Hy-Memory-30B-A3B is a 30B-A3B model. It is the part Tencent compares against Opus 5, saying it comes out ahead across major frameworks.
The framework. It handles three kinds of memory: chat, work and code.
MemHub. The user-facing piece is a CLI plus a Mac app. It syncs memory across 10+ AI tools, and Tencent frames it as the user's private memory steward.
Tencent released the SDK and MemHub on GitHub, and the model on Hugging Face under tencent. The announcement carries the tags #HyMemory and #AgentMemory.
Context
This blog covered Hy4 preview, Tencent's 770B-parameter open model, in August. Hy Memory 2.0 is a different kind of release: a 30B-A3B model and tooling focused on agent memory rather than a general-purpose flagship.
The comparison model, Claude Opus 5, launched in July.
Memory has also been a central design question for agent harnesses. VISTA gave a multimodal model a lossless visual memory it can re-inspect, while another harness solved the same benchmark with the opposite approach. Hy Memory 2.0 puts memory in a separate open layer that spans tools.