Tencent Hunyuan has upgraded Hy Memory to version 2.0 and open-sourced the memory framework, a dedicated model, and the MemHub application.
Hy Memory 1.0, released in May, was mainly built for chat memory. According to the company’s description, it can track how the same matter changes over time, identify and handle contradictions, and infer user habits that were not explicitly stated across multiple conversations. When a user changes a preference, the system keeps the old record while storing the latest decision.
Version 2.0 adds working memory and code memory
Hy Memory 2.0 adds working memory and code memory on top of the earlier chat-focused functions. Working memory records how far a task has progressed, what approaches were used, and what the final outcome was. When a similar task comes up, the system checks whether earlier experience still applies.
Code memory goes deeper by combining with Git records to trace the revision history of files and functions. Even if a piece of code was later rolled back, the system can still recover those earlier changes and show what methods had already been tried.
MemHub shares memory across AI tools
MemHub is designed to connect memory across different AI tools. Tencent said 10 tools, including Cursor, Codex, and Claude Code, can share one local memory set. After user authorization, MemHub can import historical conversations and continue syncing new records.
In Tencent’s example, if a project standard was set in Cursor, a user switching to Codex to continue development can retrieve that context directly instead of restating it.
Dedicated model and test scores
Tencent also trained a dedicated model called Hy-Memory-30B-A3B. It has 30 billion total parameters, with roughly 3 billion activated in each run. The model is designed to filter, organize, and update memory, and it can also connect to existing frameworks such as Mem0 and MemOS.
Tencent said it tested four memory systems. The average score was 55.49, slightly above Opus 5 at 54.62.