When the same task moves to a different chat window, the background often has to be explained all over again; when switching to another AI tool, already confirmed decisions may need to be explained once more. Tencent Hunyuan's open Hy Memory 2.0 attempts to provide an independent memory layer for this kind of task continuity problem.
This release covers three levels: model, framework, and application. The official repository provides the Hy Memory framework, along with the Hy-Memory-30B-A3B model dedicated to long-term memory processing, and MemHub for cross-application use.
Three types of memory for different work needs
According to the official technical description, Hy Memory distinguishes three scenarios: chat, work, and programming.
Chat memory is organized around personal information, preferences, and long-term conversation content; work memory retains task facts and reusable experience; programming memory further records the evolution trajectory of code versions, helping trace the modification process.
This division responds to a practical problem: remembering what a user likes and remembering which solution a task ultimately adopted require different information structures.
For example, after a presentation material has gone through multiple rounds of revision, the assistant needs to know which copy has been confirmed and which suggestions have been abandoned. If it retrieves only early discussions, even if that content was once accurate, it may produce errors in the current task. This is an explanation of memory requirements, not a promise of product performance.
Cross-tool memory requires clear boundaries
MemHub adopts a local-first design and can, with user authorization, aggregate conversation histories and programming records from different AI applications. The framework also provides mechanisms to filter by source and organize some memories by workspace and branch.
This means memory can become a connecting layer between multiple tools. However, local-first describes the product's data management approach; a complete deployment still requires verifying the connected models, embedding services, and their data flows.
Cross-tool sharing also needs scope control: different client projects of the same user should not mix background materials simply for retrieval convenience.
Remembering more also requires correcting old information
AIYXL observation: The quality of long-term memory cannot be judged only by how much content is recalled. More critical is whether new decisions can replace old ones, whether outdated information remains traceable, whether erroneous records can be modified or deleted, and whether clear boundaries are maintained between projects.
During enterprise trials, organizations can choose a real task that has undergone multiple rounds of revision and check whether the assistant can explain "what is currently adopted, why it is adopted, and where the basis comes from." This is closer to the requirements of daily work than simply demonstrating that it remembers a sentence.
Source: Tencent Hunyuan / Hy-Memory