Evergreen Tools
|
← Back to AI Tools

Hy-Memory 2.0: An Open Memory Stack, Model to App

On 16 October 2026 Tencent Hunyuan open-sourced the whole of Hy-Memory 2.0: Hy-Memory-30B-A3B, a model built to read and write memories with roughly 3B active parameters; a memory framework; and MemHub, a desktop app. Memory splits into chat, working and coding, covering not just what you said but how things unfolded, where a task stands and why the code changed

Tool Interface

Interactive tool will be available soon

Features

  • ✓ Three separable layers: model, framework and app each stand alone or combine into one stack
  • ✓ A dedicated memory writer: Hy-Memory-30B-A3B replaces the general-purpose LLM inside Mem0, MemOS and Hy-Memory, deciding what to retain, how to represent it and how new evidence updates old memories
  • ✓ Three memory scenes: chat memory becomes an event timeline with time, place and people; working memory tracks task progress plus the conditions under which past experience applies; coding memory tracks why code changed
  • ✓ Reported benchmarks: a 55.49 average across four memory systems, with the page listing Opus 5 at 54.04, DeepSeek V4.1 Flash at 55.47 and Qwen3-30B-A3B at 46.57
  • ✓ MemHub is a local-first desktop and CLI app where one personal memory is shared across 10 AI tools, with history import and continuous sync

How to Use

  1. Pick your layer first: if you only want an existing agent to remember things, start with the Hy-Memory SDK; swap in the 30B-A3B writer if you need a dedicated memory model
  2. For personal use, install the MemHub desktop app and connect the AI tools you already run
  3. Treat memory as sensitive and run it locally first, since the code and weights live on your own machine
  4. Test recall accuracy and forgetting behaviour on your own data before shipping; the published benchmarks are general-purpose

FAQ

When was Hy-Memory 2.0 open-sourced?

16 October 2026. Official page: https://memory.hunyuan.tencent.com/en , code at https://github.com/Tencent-Hunyuan/Hy-Memory and weights at https://huggingface.co/tencent/Hy-Memory-30B-A3B .

What changed from the previous version?

Version 1.0 in May 2026 had chat memory only, with a hierarchical structure and evolution chains, and was not open-sourced, shipping as a paper and a product. 2.0 turns chat memory into event timelines, adds working and coding memory, and open-sources model, engine and app.

Does it compete with Mem0 and MemOS or complement them?

Mostly complement. Tencent says Hy-Memory-30B-A3B is a memory-writer model designed to replace the general LLM inside those frameworks, and it plugs directly into Mem0, MemOS and Hy-Memory.

Where does the 55.49 score come from?

It is the vendor's own average across four memory systems, published alongside Opus 5 at 54.04, DeepSeek V4.1 Flash at 55.47 and Qwen3-30B-A3B at 46.57. The methodology is in the accompanying paper, so treat it as a vendor result rather than an independent verdict.

What should I watch for on privacy and licensing?

Memory holds user profiles, preferences and past conclusions, which makes it sensitive personal data. MemHub is local-first by design, so evaluate self-hosted first and give users a way to inspect, export and delete their memory; check the repository and model card for the exact licence terms.