Tencent open-sources Hy-Memory 2.0 agent memory system

Tencent released an open-source memory system for AI agents called Hy-Memory 2.0 on October 16, according to aitimes.com. The release combines a dedicated AI model, a memory management framework and a desktop application named MemHub.

MemHub connects to more than 10 AI tools through open-source agent plugins, including Claude Code, Cursor, Codex, OpenCode and OpenClaw. With user permission, it imports past conversations and work records from each tool into a single memory store and syncs newly generated records. It is offered as a command-line interface and a Mac application, with memory managed primarily on the local device.

The framework sorts memory into three types: conversation, task and code. It ships as an SDK for connecting other AI agents and as a Python package, and was published on GitHub alongside MemHub.

Model weights on Hugging Face

The dedicated model, Hy-Memory-30B-A3B, uses a Mixture-of-Experts architecture that activates about 3 billion of roughly 30 billion parameters. Its weights were published on Hugging Face.

Tencent said Hy-Memory outperformed the open-source memory frameworks Mem0, MemOS and EverOS across four benchmarks: MemoryArena, CodeMemoryBench, SWE-ContextBench and LoCoMo. Task memory averaged 55.31% on MemoryArena, while code memory reached 64.2% question-answering accuracy and 35.6% task resolution on CodeMemoryBench, plus 28.3% task resolution on SWE-ContextBench.

Conversation memory reached 72.6% accuracy on LoCoMo and cut tokens per question from 1159 to 703, a 39.3% reduction. In Tencent's technical report, the model scored a macro average of 55.49 on a 20% sample of benchmark data, ahead of Claude Opus 5 at 54.62 and DeepSeek-V4.1 Flash at 53.64.