# Tencent Hy-Memory 2.0: Open Models and Shared Memory for AI Tools

Tencent's open memory stack separates the model, framework and MemHub app. Here's what ships, what its benchmark measures, and what developers should test.

By Paras Madan | 2026-10-17


Tencent announced [Hy-Memory 2.0 on October 16, 2026](https://memory.hunyuan.tencent.com/), releasing a memory model, a developer framework and an application called MemHub. The stack aims to help AI tools retain useful context, including project knowledge that would otherwise need explaining again when switching coding assistants.

The three layers solve different parts of that problem, and developers can use them separately. A shared memory system could reduce repetition, provided it remembers the right decisions and knows when they have changed.

## What does each part do?

Hy-Memory-30B-A3B is the specialist model. It decides what information to save, how to represent it and how new evidence should update a stored memory. Its mixture-of-experts architecture has about 30 billion total parameters, with roughly 3 billion active during inference.

The Hy-Memory framework manages the surrounding process. Its chat pipeline handles conversational memory, while working memory tracks task facts and reusable experience. Coding memory follows changes through a task's history.

MemHub brings that system to existing AI applications. Tencent lists ten supported tools, including Cursor, Codex and Claude Code. It imports authorized conversation history and makes relevant memory available across connected tools.

That separation helps when choosing where to start. Someone building an agent can evaluate the framework. A team already using a memory framework could test the specialist writer. MemHub addresses people who want continuity between applications they already use.

## Did a 3B model beat Claude?

The [model card](https://huggingface.co/tencent/Hy-Memory-30B-A3B) reports a 55.49 macro-average score against Claude Opus 5's 54.62 across four memory frameworks. Calling that a general win over Claude would stretch the result considerably.

The experiment replaces the memory writer while holding the retrieval configuration, answer model and other components fixed. Each framework's score averages six metrics, and the main comparison uses fixed 20% benchmark subsamples. EverMemOS was held out from training.

The [technical paper](https://arxiv.org/abs/2610.22196), submitted October 15, describes training that rewards memories for helping answer later questions. These are Tencent's reported results, and they support evaluating a specialized writer for that role. They don't establish a general coding or reasoning advantage. Also, the 3B figure describes active parameters; deployment still has to account for the larger model.

A useful test would include decisions that changed halfway through a project. Can the agent recover the current convention without reviving an abandoned approach? A memory that preserves a wrong decision can create more work than starting fresh.

## What can developers use today?

The [framework repository](https://github.com/Tencent-Hunyuan/Hy-Memory) documents the public Python package, hy-memory 2.0.2. Both framework and model carry Apache 2.0 licences. The SDK needs an LLM endpoint and an embedding endpoint, with OpenAI-compatible services supported. Serving infrastructure or provider usage still has a cost.

The [MemHub documentation](https://github.com/Tencent-Hunyuan/Hy-Memory/blob/main/memhub/README.md) describes a local-first application with separate authorization for reading source data, model processing and access by connected tools. Linux and macOS are release targets; Windows is unsupported. Check the published release assets before assuming a packaged installer exists for your machine.

Local storage also leaves a separate question: where does the configured model process that information? Teams should inspect those endpoints before importing work conversations or code traces.

For a first evaluation, use one repository and a small set of deliberately changing instructions. Measure whether later tasks follow the latest decision, how much correction they need and what the extra memory processing costs. That gives a much clearer adoption signal than the headline benchmark alone.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain their models, APIs and engineering updates in stories that developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).