News from the IT community on October 16th: Tencent Hunyuan officially announced the release of the Agent Memory Framework Hy Memory version 2.0 today, and made the dedicated models, memory framework, and MemHub application capabilities open source.

The three-layer capabilities of Hy Memory 2.0 can be used independently or combined to form a complete Memory solution. Tencent states that currently, Hy Memory has been utilized in various business areas within Tencent, including internal finance, information management, 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 memories: chat, work, and code. Developers can integrate it with Agent through SDK.

Tencent stated that this framework is the first in the industry to propose a memory framework that integrates working memory and code memory. In four scenarios involving chatting, working, and coding, this framework outperforms other open-source frameworks such as Mem0, MemOS, EverOS, OpenViking, RepoMemory, and Simple RAG.

Specifically, working memory contributed an average of 55.31% to the travel planning and combined shopping tasks in MemoryArena, which is 10.2 percentage points ahead of the second-best performer, EverOS; code memory achieved a code answering accuracy of 64.2% and a code task resolution rate of 35.6% in CodeMemoryBench, and a code task resolution rate of 28.3% in SWE-ContextBench, which is higher than the 25.3% in MemOS and Mem0; chat memory had an accuracy rate of 72.6% in LoCoMo, which is higher than the 71.9% in EverOS, and the number of Token used per question decreased from 1,159 to 703, representing a reduction of 39.3%.

In terms of chat memory, Hy Memory 2.0 has evolved from processing scattered facts to forming a complete narrative of events. It is capable of organizing dispersed information across different times and conversations into a coherent story that includes key elements such as time, location, and participants. When faced with questions that require combining multiple clues to answer, Hy Memory first merges duplicate information, resolves any inconsistencies, and then passes the clear and coherent memories to the model for processing.

Working memory and code memory are new capabilities added in version 2.0, which allow for the memorization of tasks across multiple dimensions. All information is updated over time, including task objectives, established methods, stage results, as well as reusable experiences and applicable conditions. Tencent states that when reusing experience, Hy Memory first determines whether the current situation meets the applicable conditions for that experience, and only then calls upon it precisely as a form of assistance. For long-term tasks, before each continuation, Hy Memory first identifies what is currently being worked on, then retrieves the relevant decisions, still-applicable experiences, and the historical work process.

At the model level, Hy-Memory-30A3B is specifically trained for memory-related tasks and has approximately 3 billion activation parameters. According to the paper information on arXiv, this model was trained on three different frameworks: Mem0, MemOS, and Hy-Memory. It ranked first in all four evaluation frameworks that used a fixed 20% subset of the baseline data, with an average macro score of 55.49%, which is higher than Claude Opus's score of 54.62%.

In the tests involving the integration of four memory systems: Hy Memory, Mem0, MemOS, and EverMemOS, Hy-Memory-30A3B scored 55.49, which is higher than Opus's score of 54.62, DeepSeek V4's 53.64, Flash's 53.64, and Qwen3-30B-A3B's 44.37.

At the application level, MemHub is a local memory management application that offers two usage methods: CLI and Mac. Once users authorize the associated applications, MemHub can import the historical sessions from various tools and consolidate them into local memory through the built-in Hy Memory SDK. Subsequently, new sessions generated by these tools can also be synchronized in real-time to the same local memory.

MemHub supports one-click integration with AI tools via plugins. When users initiate tasks within any associated tool, MemHub allows them to manage and utilize relevant memories as contextual supplements. Currently, it covers 10 types of AI tools, including WorkBuddy, Cursor, Codex, Pi, Qwen Code, Hermes, DeepSeek Harness, and OpenCode, encompassing both commercial programming tools and open-source Agent.

Currently, Hy Memory SDK and MemHub have been open-sourced on GitHub, and the model weights for Hy-Memory-30B-A3B have been released on Hugging Face.

Tencent Hunyuan stated that it will continue to collaborate deeply with more businesses to refine memory capabilities in real-world scenarios, enabling Agent to maintain continuity across longer tasks and with a greater variety of tools.

Reference materials:

• Hy Memory SDK & MemHub : https :// github.com / Tencent-Hunyuan / Hy-Memory
• Hy-Memory-30B-A3B Model weight download: https :// huggingface.co / tencent / Hy-Memory-30B-A3B