Tencent Hunyuan Releases Open-Source Agent Memory Framework Hy Memory 2.0

Tencent Hunyuan announced on October 16 the release and open-sourcing of Hy Memory 2.0, an agent memory framework that provides three layers: a dedicated memory model, the framework itself, and the MemHub application. The 2.0 version adds working memory and code memory, which the company says makes it the first memory framework to support both types.

The framework layer handles storage, updating, and evolution of chat, working, and code memories, allowing developers to integrate it into agents via an SDK. According to benchmarks, Hy Memory 2.0 outperformed open-source frameworks including Mem0, MemOS, and EverOS across four benchmarks. On MemoryArena's travel planning and combined shopping tasks, working memory scored 55.31%, 10.2 percentage points higher than EverOS. Code memory achieved 64.2% accuracy on CodeMemoryBench question answering and a 35.6% task resolution rate, and 28.3% on SWE-ContextBench. Chat memory scored 72.6% on LoCoMo, with token usage per question dropping 39.3% from 1,159 to 703.

On the model side, Hy-Memory-30A3B is specialized for memory scenarios with about 3B activated parameters. On a fixed 20% benchmark subset across four evaluation frameworks, it scored an average of 55.49%, surpassing Claude Opus 5's 54.62%.

MemHub, the application layer, is a local memory management tool available via CLI and Mac app. After user authorization, it imports historical conversations from connected tools and consolidates them into local memory. It supports plugin-based integration with AI tools, covering 10 tools such as WorkBuddy, Cursor, Codex, and Qwen Code.

The Hy Memory SDK and MemHub are now open-sourced on GitHub, and the Hy-Memory-30B-A3B model weights are available on Hugging Face.