Key Takeaways
- Hy Memory 2.0 is an open-source memory system from Tencent Hunyuan made of three parts: the Hy-Memory-30B-A3B model, a memory framework, and the MemHub app.
- The framework keeps chat, work and code memories separate, while the roughly 3B-active-parameter model retrieves the relevant context when you need it.
- MemHub ships as a CLI and a Mac app, imports permitted history, and syncs memories across more than ten AI tools, including Cursor.
- Tencent claims the model beats Opus 5 on major memory benchmarks and ranks first open-source on four benchmarks; treat those numbers as vendor claims and verify them yourself.
Quick Facts
| Type | Open-source memory model, framework and app |
|---|---|
| Devices | Mac app plus CLI; check the repo for other platforms |
| Price | Open source; check the official repository for terms |
| Languages | Not confirmed in the notes; check the official site |
| Highlight | One memory layer syncing 10+ AI tools |
| Learning curve | Moderate; start with MemHub docs and a small test |
| Category | Dev & Web Tools |
| Creator | Tencent Hy |
| Official site | github.com ↗ |
| Heat | 🔥 155.9K Heat |
| Last updated |
What Tencent Hunyuan Built and Why Memory Matters
Every AI assistant forgets you the moment a session ends. You re-explain your stack, your naming conventions and your preferences in each new chat, and that context dies when you close the tab. Hy Memory 2.0, announced by Tencent Hunyuan on October 16, 2026, attacks that problem directly. It is an open-source memory system made of three pieces: a dedicated memory model, a framework that organizes what gets remembered, and an app that moves those memories between tools.
The model is Hy-Memory-30B-A3B, a mixture-of-experts design with roughly 3 billion active parameters. The framework splits memory into three domains, chat, work and code, so a casual conversation does not pollute your repository context. MemHub is the delivery layer: a command-line tool and a Mac app that imports permitted history and syncs memories across more than ten AI tools. Tencent describes the whole package as a private memory steward.
The timing explains the attention. Developers now juggle several AI tools daily, and each one starts from zero. A single memory layer that follows you between them is genuinely useful, and Tencent's claim that Hy-Memory-30B-A3B beats Opus 5 on major memory benchmarks, ranking first among open-source systems on four of them, gives people a concrete reason to look. Those are vendor claims, so verify them against your own workloads.
- Three parts, one system
The model retrieves, the framework organizes, and MemHub transports. You can adopt them together or study each piece separately in the repository.
- Separated memory domains
Chat, work and code memories live apart, which keeps personal context out of professional retrieval and vice versa.
- Cross-tool sync
MemHub is designed to carry memories across more than ten AI tools, with Cursor cited as an example, so context survives when you switch editors.
- Open source
Source code and setup instructions are published in the official repository on github.com, so you can inspect exactly what is stored and how.
How the Model, Framework and MemHub Fit Together
Think of the framework as the librarian and the model as the search engine. When you interact with a connected tool, the framework decides which domain the interaction belongs to, extracts durable facts, and stores them in a structured memory. Later, when a new request arrives, the framework queries the model, and Hy-Memory-30B-A3B retrieves the pieces of context that matter. Because only about 3 billion parameters are active at a time, retrieval stays fast despite the larger total parameter count.
MemHub is what makes this portable. It runs as a CLI and as a Mac app, imports history you explicitly permit, and syncs the resulting memory store across your AI tools. That means a preference you teach one assistant can inform another, and a code convention you establish in one editor can follow you into the next. The notes describe this as syncing across more than ten AI tools, with Cursor named as an example; the full supported list is worth checking on the official site.
The privacy framing matters here. Tencent calls it a private memory steward, and the architecture supports that story: memories are organized locally around your domains rather than dumped into one undifferentiated pool. Still, memory systems by definition retain sensitive material, so read the repository documentation carefully before letting MemHub ingest chats, API keys or customer data.
- Retrieval, not replay
The model selects relevant memories instead of stuffing your entire history into every prompt, which keeps context windows manageable.
- Domain routing
The framework classifies each interaction as chat, work or code, so retrieval draws from the right pool for the task at hand.
- Import with consent
MemHub imports permitted history, meaning you choose what it ingests rather than having it scrape everything automatically.
- Verify the benchmark claims
Tencent says the model beats Opus 5 on major memory benchmarks; reproduce the tests on your own data before trusting the ranking.
Setting Up a Safe First Workflow
The sensible way to adopt Hy Memory 2.0 is small and boring. Start with the documentation in the official repository on github.com, install MemHub, and point it at a single non-sensitive project. A toy repository or a personal notes folder works well. Watch what the framework chooses to remember, how it labels the domain, and whether retrieval actually improves your results. Only after that should you consider connecting real work.
Once the basics behave, expand deliberately. Add one AI tool at a time and confirm that memories transfer as expected before adding the next. Keep the chat domain isolated from the code domain until you understand how the framework classifies borderline cases, since a misrouted memory can surface personal context in a professional prompt. If something looks wrong, the open-source code lets you trace the decision rather than guess.
Finally, treat memory as data you own and audit. Periodically review what has been stored, delete anything stale, and keep credentials out of the system entirely. The notes specifically warn against importing private chats, API keys or customer data during early testing, and that advice holds even after you are comfortable. A memory layer is only as trustworthy as the hygiene you apply to it.
- Test on a throwaway project
A small, non-sensitive repository reveals how the framework stores and retrieves without risking anything you care about.
- Add tools one at a time
Connecting a single AI tool first makes it obvious whether sync works before complexity hides the problem.
- Keep secrets out
Do not let MemHub ingest API keys or customer data; the notes flag this explicitly for early testing.
- Audit stored memories
Review and prune the memory store regularly so outdated or wrong facts do not quietly steer future answers.
How It Compares and Who Should Bother
Most AI tools now offer some form of built-in memory, but those stores are siloed inside each product. Hy Memory 2.0's pitch is the opposite: one layer you control that spans more than ten tools, including Cursor, with an open model and inspectable code underneath. That combination of portability and transparency is the differentiator, not raw model size, since the active parameter count is modest by frontier standards.
The trade-offs are real. You are running and maintaining a memory system yourself, which means setup effort, ongoing auditing and a dependency on Tencent's framework updates. Vendor benchmark claims about beating Opus 5 should be treated as a starting point for your own evaluation, not a conclusion. And because the notes do not confirm platform coverage beyond the CLI and Mac app, Windows and Linux users should check the official site before planning an adoption.
Who benefits most? Developers who work across several AI coding tools and lose context constantly, teams that want memory to stay under their own control, and anyone curious about how retrieval-based memory systems work internally. If you use a single assistant and rarely switch, built-in memory may be enough. If you live in a multi-tool workflow, Hy Memory 2.0 is worth a careful trial.
- Portability versus built-in memory
Native memory keeps you inside one product; Hy Memory 2.0 follows you across tools, which matters only if you actually switch.
- Open versus closed
The open-source model and framework let you inspect storage and retrieval logic, something closed memory features do not offer.
- Effort versus convenience
Self-hosting means setup and maintenance work; weigh that against the convenience of a vendor-managed memory.
- Check platform support
The notes confirm a CLI and Mac app; confirm availability for your operating system on the official site before committing.
Step-by-Step Guide
Read the repository documentation first
Go to the official repository on github.com and read the setup instructions for the model, framework and MemHub. Understand what gets stored, where it lives and which tools are supported before installing anything. This step prevents surprises later.
Install MemHub on your Mac or via CLI
Follow the repository instructions to install MemHub as either the Mac app or the command-line tool. Pick the interface you will actually use day to day. Confirm the installation completes cleanly before connecting any AI tools.
Connect one AI tool and import limited history
Link a single supported tool, such as Cursor, and let MemHub import only history you explicitly permit. Use a small, non-sensitive project for this first connection. Verify that memories appear and are retrievable inside that tool.
Test retrieval quality on real tasks
Ask the connected assistant questions that depend on remembered context, like your preferred naming style or project structure. Check whether the framework routes the interaction to the correct chat, work or code domain. Note any misrouting so you can adjust how you work.
Expand carefully and audit regularly
Add more AI tools one at a time once sync proves reliable, and periodically review the stored memories for stale or sensitive entries. Keep API keys and customer data out of the system. Re-check the official site for updates and platform support.
Frequently Asked Questions
Is Hy Memory 2.0 free to use?
The source code and instructions are published in the official repository on github.com, and the project is described as open source, so you can inspect and run it yourself. The notes do not state specific licensing terms or any paid tiers, so check the repository and official site for the exact license and any commercial conditions before deploying it in a business setting.
Which devices and platforms does MemHub support?
The notes describe MemHub as available as a CLI and a Mac app, and say it syncs memories across more than ten AI tools, with Cursor named as an example. Support for Windows, Linux or mobile is not confirmed in the notes. Check the official site and repository for the current platform list before planning your setup.
How does the memory system actually work?
The framework separates memories into chat, work and code domains, extracting durable facts from your interactions and storing them in a structured form. When a new request arrives, the Hy-Memory-30B-A3B model retrieves the relevant memories, using roughly 3 billion active parameters for speed. MemHub then carries that memory store between your connected AI tools.
Does Hy-Memory-30B-A3B really beat Opus 5?
Tencent claims the model beats Opus 5 on major memory benchmarks and ranks first among open-source systems on four benchmarks. These are vendor claims made at announcement, not independently verified results. Because the model and framework are open source, you can reproduce the benchmarks or test retrieval quality on your own data before drawing conclusions.
Is it safe to import my chats and code history?
The notes advise starting with a small, non-sensitive project and explicitly warn against importing private chats, API keys or customer data during early testing. Memory systems retain whatever they ingest, so treat setup as a data-handling decision. Read the repository documentation, import only what you permit, and audit stored memories regularly.
