IT Home, October 16: Tencent Hunyuan officially announced today the launch of the Agent Memory Framework Hy Memory 2.0, open-sourcing three layers of capabilities: the dedicated model, the memory framework, and the MemHub application.
Hy Memory 2.0’s three-layer capabilities can be used independently or combined into a complete Memory solution. Tencent states that Hy Memory is currently being used across multiple business areas within the company, including finance, news, knowledge management, office agents, social entertainment, and gaming.
According to Tencent, the framework layer of Hy Memory 2.0 is responsible for storing, updating, and evolving memories for chat, work, and code, and developers can integrate it into agents via the SDK.
Tencent stated that this framework is the first in the industry to propose a memory architecture combining working memory and code memory. On four benchmarks covering chat, work, and code scenarios, Hy Memory outperforms open-source frameworks such as Mem0, MemOS, EverOS, OpenViking, RepoMemory, and Simple RAG, achieving top results.
Specifically, working memory achieved an average score of 55.31% on MemoryArena’s travel planning and combined shopping tasks, outperforming the second-best EverOS by 10.2 percentage points; code memory attained a code question-answering accuracy of 64.2% and a code task resolution rate of 35.6% on CodeMemoryBench, and a code task resolution rate of 28.3% on SWE-ContextBench, surpassing MemOS and Mem0’s 25.3%; chat memory reached an accuracy of 72.6% on LoCoMo, higher than EverOS’s 71.9%, while reducing the average tokens per question from 1,159 to 703, a 39.3% decrease.
In terms of chat memory, Hy Memory 2.0 has evolved from fragmented facts to comprehensive event narratives, enabling it to organize scattered content across time and conversations by identifying key elements such as time, location, and participants around each event. When faced with questions requiring the integration of multiple clues, Hy Memory first consolidates duplicate information, resolves contradictions, and then delivers a clear, coherent memory to the model.
Working memory and code memory are new capabilities introduced in version 2.0, utilizing multiple dimensions to remember a task, organizing all information updates over time, including task objectives, established approaches and phase outcomes, as well as reusable experiences and their applicable conditions. Tencent states that when reusing experiences, Hy Memory first determines whether the current context meets the applicable conditions for that experience, and only invokes the relevant experience precisely when those conditions are satisfied. For long-term tasks, before resuming each task, Hy Memory first identifies what is currently being addressed, then retrieves related decisions, still-applicable experiences, and historical workflow details.
At the model level, Hy-Memory-30A3B is specifically trained for memory tasks, with approximately 3B activated parameters. According to arXiv paper information, this model was trained on the Mem0, MemOS, and Hy-Memory frameworks and achieved first place across all four evaluation benchmarks using a fixed 20% baseline subset, with a macro-average score of 55.49%, surpassing Claude Opus 5’s 54.62%.
In tests involving the Hy Memory, Mem0, MemOS, and EverMemOS memory systems, Hy-Memory-30A3B scored 55.49, outperforming Opus 5 at 54.62, DeepSeek V4.1 Flash at 53.64, and Qwen3-30B-A3B at 44.37.
At the application level, MemHub is a local memory management tool available via CLI and a Mac app. After users authorize connected applications, MemHub imports historical sessions from various tools and stores them as local memory using its built-in Hy Memory SDK. New sessions generated by these tools are then synchronized in real time into the same memory repository.
MemHub supports one-click integration with AI tools via plugins. When users initiate tasks within any connected tool, MemHub can manage and utilize relevant memories as contextual supplements. It currently supports 10 AI tools, including WorkBuddy, Cursor, Codex, Pi, Qwen Code, Hermes, DeepSeek Harness, and OpenCode, covering both commercial programming tools and open-source agents.
The Hy Memory SDK and MemHub are now open-sourced on GitHub, and the Hy-Memory-30B-A3B model weights have been released on Hugging Face.
Tencent Hunyuan stated that it will continue to deeply collaborate with more business areas, refining its memory capabilities in real-world scenarios to enable agents to maintain continuity across longer tasks and a greater number of tools.
Reference materials:
- Hy Memory SDK & MemHub: https://github.com/Tencent-Hunyuan/Hy-Memory
- Download Hy-Memory-30B-A3B model weights: https://huggingface.co/tencent/Hy-Memory-30B-A3B