Tencent’s Hunyuan team has open-sourced Hy Memory 2.0, an artificial intelligence memory system designed to help AI assistants retain useful information across conversations, work sessions and coding tasks. Released on October 16, 2026, the system combines a specialised memory model, a software development framework and MemHub, an application that brings authorised conversation histories from multiple AI tools into a shared memory system. Tencent is releasing these components under the Apache 2.0 licence, making the project available for developers to inspect, use and adapt under the licence’s terms.
The launch addresses a growing limitation of AI assistants: context often remains trapped inside individual applications. A developer might explain a project’s architecture to one coding agent, establish preferences in another assistant and document a decision in a third tool, only to repeat that information when switching between them. Hy Memory 2.0 aims to reduce this repetition by preserving relevant facts, task progress and the reasoning behind code changes. Tencent says MemHub can connect 10 AI tools, although users must authorise the relevant integrations and the actual experience depends on supported applications and configuration.
Key Takeaways
- Open-source release: Tencent Hunyuan released Hy Memory 2.0 on October 16, 2026, covering its model, framework and application.
- Three memory types: The system supports chat memory, working memory and coding memory.
- Shared context: MemHub is designed to synchronise authorised conversation histories and memory across 10 AI tools.
- Dedicated model: Hy-Memory-30B-A3B is a specialised memory model with approximately 30 billion total parameters and a smaller number activated per inference.
- Open licence: Tencent has released the public code and model assets under Apache 2.0 terms.
- Benchmark claims: Tencent reports a 72.6% result on the LoCoMo conversational-memory benchmark and leading results for its framework on several comparisons.
- Privacy considerations: MemHub is designed as a local-first application, but information sent to third-party AI services remains subject to the permissions and data-handling arrangements involved.
What Is Tencent Hy Memory 2.0?
Hy Memory 2.0 is a memory layer for AI systems, rather than a general-purpose chatbot intended to replace models such as Claude, ChatGPT or other assistants.
A conventional AI conversation can use the messages available in its current context. But when a conversation becomes long, an application changes or a new task begins, important details may be unavailable. Even when a tool supports persistent memory, that memory may not automatically transfer to another application.
Hy Memory attempts to organise useful information so an AI agent can retrieve relevant details when needed. Instead of simply storing a large collection of old messages, the framework aims to identify relationships between facts, organise task progress and retrieve context suited to the current request.
Version 2.0 expands the project beyond conversational memory. It adds dedicated working and coding memory, making it more relevant to people who use AI assistants for ongoing projects rather than isolated questions.
The release contains three connected components:
| Component | Purpose |
|---|---|
| Hy-Memory-30B-A3B | A specialised model trained for memory-related operations |
| Hy Memory framework | A software development kit and memory engine developers can integrate into AI agents |
| MemHub | An application designed to share authorised memory across supported AI tools |
Each component can be used independently or combined into a broader memory system. This modular approach gives developers a choice between adopting the complete solution and incorporating only the parts they need.
Three Types of Memory: Chat, Work and Coding
Hy Memory 2.0 separates memory into three categories because different tasks require different information to be retained.
1. Chat memory
Chat memory is designed to preserve information from conversations over time. It attempts to organise scattered facts into connected events and relationships, helping an AI system recall what happened, when it happened and how later developments changed the situation.
For example, a user might explain a preference in one conversation and revise it several weeks later. A useful memory system needs to retain the relevant preference while recognising that older information may have become outdated.
This is more complicated than searching for a matching keyword. The system must identify relevant details, reduce duplication and handle conflicting information before supplying context to the model.
2. Working memory
Working memory focuses on ongoing tasks and reusable experience. It can preserve a task’s objective, decisions already made, current progress and lessons that may be useful later.
Consider an AI assistant helping plan a software migration. It may need to remember which approach was selected, which components have already been migrated and which issues remain unresolved.
Without persistent working memory, a user may need to reconstruct the task every time a new session begins. Hy Memory aims to make that transition smoother by retrieving the relevant task state and previously established decisions.
The system also aims to distinguish reusable experience from rules that apply only under particular conditions. That distinction matters because a solution that worked in one project may be inappropriate in another.
3. Coding memory
Coding memory is designed for software development tasks that span multiple sessions. It focuses on code changes, the decisions behind those changes and the history needed to continue work without starting from scratch.
An AI coding assistant may remember the latest code but fail to retain why a developer rejected an alternative design. That can lead to repeated suggestions, duplicated debugging or changes that reverse earlier decisions.
Hy Memory’s coding memory aims to preserve relevant history and retrieve it when the agent resumes a task. This could be useful for maintaining large codebases, implementing features over several days or coordinating work across different coding tools.
However, memory does not replace source control, tests, documentation or code review. Developers still need those systems to verify what changed and whether the change is correct.
MemHub Connects Memory Across 10 AI Tools
MemHub is the application layer of Hy Memory 2.0. Its main purpose is to create a shared memory store from conversations and coding traces that users choose to import from supported AI applications.
Tencent’s materials describe a workflow in which users authorise their applications, import previous sessions and allow new conversations to be added to the shared memory system. The relevant memory can then be retrieved by another connected tool when a task requires it.
The project lists integrations involving tools such as Cursor, Codex, Claude Code, OpenCode, Qwen Code and other AI coding or assistant environments.
This matters because AI tools increasingly specialise in different tasks. A developer may use one assistant to plan a feature, another to edit code and a third to investigate bugs. Without a shared context layer, each tool can behave as though the work began from scratch.
MemHub is intended to reduce that fragmentation.
| Without shared memory | With a shared memory layer |
|---|---|
| Preferences may need to be repeated in different tools | Relevant preferences can be retrieved across authorised tools |
| Project decisions may remain in separate conversations | Important decisions can be collected into shared memory |
| New sessions may require manual explanations | The system can retrieve relevant task history |
| Coding agents may repeat previous investigations | Historical decisions and experience can help guide the next step |
This comparison describes the intended workflow, not a guarantee that every detail will transfer perfectly between applications.
Local-first design and platform availability
Tencent describes MemHub as a local-first application. Its published documentation says imported memory is managed in a local data directory, allowing users to maintain a shared memory layer without making a single cloud service the only repository.
The project’s published documentation identifies macOS and Linux as supported targets for the initial application releases. Users should check the current repository and release notes for the latest platform support, because availability can change as the project develops.
Local storage can give users more control over where memory is kept, but it does not automatically mean every interaction stays on the device. When a task uses an external AI model or service, the relevant information may need to be transmitted to that provider.
Users should therefore review the permissions, model configuration and data-handling terms before connecting sensitive work accounts or private conversations.
Hy-Memory-30B-A3B: A Model Built for Memory Tasks
The model component, Hy-Memory-30B-A3B, is designed to perform memory-related operations rather than act as a general-purpose assistant for every task.
The name indicates approximately 30 billion total parameters, with around three billion activated per inference under the model’s architecture. The distinction matters because total parameter count and active parameter count describe different aspects of a model’s size and computation.
A specialised memory model can be useful because storing information and retrieving the right information are different problems from generating a general answer. The system must decide which facts are worth retaining, how they relate to earlier details and which information is relevant to a later request.
Tencent says the model can be used alongside Hy Memory or integrated into other memory systems, including frameworks such as Mem0 and MemOS. That flexibility could help developers experiment with different memory architectures without having to rebuild their entire AI application.
As with other open model releases, practical performance will depend on hardware, configuration, the underlying language model and the quality of the information being stored.
What Do Tencent’s Benchmark Results Show?
Tencent reports that Hy Memory 2.0 achieved a 72.6% result on the LoCoMo benchmark, which evaluates an AI system’s ability to answer questions about long conversations and retrieve relevant information.
The company also reports leading results for its memory framework across four benchmarks covering chat, working and coding scenarios.
These claims are relevant because long-term memory systems can be difficult to compare. A system may perform well at remembering personal facts but struggle to track the progress of a complex software task. Testing several memory scenarios can provide a broader picture of its capabilities.
However, benchmark results should be interpreted carefully. Scores can depend on the dataset, the underlying language model, the evaluation procedure and the way questions are selected. Results reported by the developer should not automatically be treated as independent confirmation of superiority in every real-world application.
For developers, the most useful test may be whether the system reliably remembers the information their own tasks require, avoids retrieving outdated details and reduces the amount of repeated prompting needed.
Why Open-Sourcing the Entire Stack Matters
Tencent’s decision to release the model, framework and application gives developers several ways to adopt the technology.
A team that already operates an AI agent could integrate the SDK without adopting MemHub. Another developer could test the specialised model with an existing memory framework. Individual users may prefer the application because it offers a more direct way to connect supported tools.
The Apache 2.0 licence is also significant. It generally permits commercial use, modification and redistribution subject to its conditions, including applicable notices and licence obligations. Developers should review the licence files for the exact components they intend to use.
An open release can encourage experimentation and independent evaluation. Developers may identify weaknesses, add integrations or adapt the memory architecture for specific industries.
The model’s usefulness will ultimately depend on more than its availability. Documentation, stable releases, integrations, performance and privacy controls will determine whether developers adopt it for production systems.
The Bigger Picture
Hy Memory 2.0 reflects a broader shift in AI development from assistants that answer isolated prompts toward agents expected to maintain context over longer periods. As people use more specialised tools for coding, research and productivity, the ability to preserve decisions and retrieve relevant information across applications could become an important part of the AI software stack.
Tencent’s open-source release gives developers a model, framework and application to explore that problem. Its reported benchmark results are promising, but practical adoption will depend on independent testing, reliable integrations and clear data controls. Shared memory can reduce repeated explanations, yet it also raises questions about what gets stored, how outdated information is handled and when private context is transmitted to external services.
Looking Ahead
Developers can explore Hy Memory’s source code, model assets and documentation through Tencent Hunyuan’s public repositories. The immediate next step for teams evaluating the system will be to test whether it preserves useful context in their own workflows and whether its integrations fit their preferred AI tools. Users should also verify supported platforms and review data permissions before importing conversation histories.
Over time, persistent memory could become a standard feature of AI agents, much as databases and version control are standard parts of software development. Hy Memory 2.0 is one approach to that future, combining specialised memory operations with a shared application layer. Whether it gains wider adoption will depend on its reliability, ease of integration and ability to offer meaningful benefits without compromising privacy or control.
Frequently Asked Questions
1. What is Tencent Hy Memory 2.0?
Hy Memory 2.0 is an open-source AI memory system from Tencent Hunyuan. It combines a specialised memory model, a developer framework and MemHub, an application designed to share authorised memory across supported AI tools.
2. What is MemHub used for?
MemHub imports and organises authorised conversation histories from supported AI applications so relevant context can be retrieved across connected tools. Its purpose is to reduce repeated explanations and help AI agents continue ongoing work.
3. Which AI tools can MemHub connect?
Tencent lists integrations for tools including Cursor, Codex, Claude Code, OpenCode and Qwen Code, among others. Users should check the latest repository documentation for the current list and support status.
4. Is Hy Memory 2.0 open source?
Yes. Tencent has released the public code and model assets under Apache 2.0 terms. Developers should review the relevant licence and documentation for each component they use.
5. Does Hy Memory 2.0 keep all data on the user’s computer?
MemHub is described as local-first and stores its memory data locally. However, when a connected task uses an external AI model or service, authorised content may be sent to that provider. Users should review permissions and data-handling details.
6. Does Hy Memory 2.0 replace ChatGPT or other AI models?
No. It is a memory layer that can work with AI agents and models. Its purpose is to help those systems retain and retrieve useful context rather than replace a general-purpose assistant.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
