IT Home On 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 specialized model, the memory framework, and the MemHub application.
Hy Memory 2.0's three layers of capabilities can be used independently or combined into a complete Memory solution. Hy Memory is currently used in multiple Tencent internal businesses across finance, news, knowledge management, office Agent, social entertainment, and gaming.
According to Tencent, the framework layer of Hy Memory 2.0 is responsible for the storage, updating, and evolution of three types of memory: chat, work, and code, and developers can integrate it into Agents via the SDK.
Tencent stated that this framework is the industry's first memory framework to propose working memory and code memory. Across 4 Benchmarks in the three scenario categories of chat, work, and code, Hy Memory ranked first compared with open-source frameworks such as Mem0, MemOS, EverOS, OpenViking, RepoMemory, and Simple RAG. Specifically, working memory achieved an average score of 55.31% on MemoryArena's travel planning and combined shopping tasks, leading the runner-up EverOS by 10.2 percentage points; code memory achieved 64.2% code Q&A accuracy and 35.6% code task resolution rate on CodeMemoryBench, and a 28.3% code task resolution rate on SWE-ContextBench, higher than the 25.3% of MemOS and Mem0; chat memory achieved 72.6% accuracy on LoCoMo, higher than EverOS's 71.9%, while the Tokens used per question dropped from 1,159 to 703, a reduction of 39.3%.
For chat memory, Hy Memory 2.0 upgrades from scattered facts to complete event threads, able to organize scattered content across time and conversations around events as the main subject, sorting out key elements such as time, place, and people. When encountering questions that require combining multiple clues to answer, Hy Memory first merges duplicate information and strips out contradictions, then hands the clear memory to the model. Working memory and code memory are new capabilities in the 2.0 version, memorizing a task across multiple dimensions, organizing all information updates chronologically, including task goals, established approaches and stage results, as well as reusable experience and its applicable conditions. When reusing experience, Hy Memory first judges whether the current situation matches the experience's applicable conditions, and only applies the corresponding experience as assistance when it does. For long-running tasks, each time before resuming a task, Hy Memory first judges which matter is currently being handled, then retrieves relevant decisions, still-applicable experience, and the historical work process.
At the model level, Hy-Memory-30A3B is specifically trained for memory scenarios, with about 3B activated parameters. According to information from the arXiv paper, the model was trained on the three frameworks Mem0, MemOS, and Hy-Memory, and ranked first on all four evaluation frameworks using a fixed 20% benchmark subsample, with a macro average score of 55.49%, higher than Claude Opus 5's 54.62%.
In testing across the four memory systems Hy Memory, Mem0, MemOS, and EverMemOS, Hy-Memory-30A3B scored 55.49, higher than Opus5's 54.62, DeepSeek V4.1 Flash's 53.64, and Qwen3-30B-A3B's 44.37.
At the application level, MemHub is a local memory management application, available in two forms: a CLI and a Mac APP. After users authorize linked applications, MemHub can import historical sessions from each tool and, through the built-in Hy Memory SDK, consolidate them into local memory; new sessions generated by each tool thereafter can be synchronized in real time into the same memory.
MemHub supports one-click integration with AI tools in the form of plugins; when a user initiates a task in any linked tool, MemHub can manage and use relevant memory as contextual supplement. It currently covers 10 AI tools including WorkBuddy, Cursor, Codex, Pi, Qwen Code, Hermes, DeepSeek Harness, and OpenCode, spanning commercial programming tools and open-source Agents.
Currently, the Hy Memory SDK and MemHub have been 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 collaborate deeply with more businesses, polishing memory capabilities in real-world scenarios, so that Agents can maintain continuity across longer tasks and more tools.
References:
Hy Memory SDK&MemHub: https://github.com/Tencent-Hunyuan/Hy-Memory
Hy-Memory-30B-A3B model weights download: https://huggingface.co/tencent/Hy-Memory-30B-A3B
Related reading: