Hy-Memory-30B-A3B + Hy Memory: Local Memory Writer for AI Agents

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Hy-Memory-30B-A3B + Hy Memory: Local Memory Writer for AI Agents · Channel: Fahd Mirza · 07:52 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Local and open-source models

Summary

The video demonstrates installing Tencent Hunyuan’s Hy-Memory-30B-A3B memory-writing model and the Hy Memory SDK, then connecting them to a local vLLM server. In a test, the system stores that Alex lives in Seattle, processes an update that he moved to Sydney, and replaces the outdated memory. The presenter also describes the project’s open-source license, benchmark results, and two-stage training setup. He concludes that he plans to use the tool in production.

Key points

Hy-Memory-30B-A3B is presented as a model trained to decide what memories to keep and how to update them.
Hy Memory is Tencent’s memory engine and SDK; the model acts as the memory writer.
The demo uses a local vLLM server running on an Nvidia H100 with 80GB of VRAM.
The SDK’s Pro mode uses an LLM to extract and reconcile memories.
An embedding model converts stored memories into vectors; the presenter uses Alibaba’s API for the demonstration.
The test adds Alex’s Seattle location and then updates it to Sydney, checking whether the old fact is superseded.
The demo uses SQLite behind the scenes and shows the updated profile stating that Alex lives in Sydney.
The presenter reports a benchmark score of 55.49 for Hy Memory, ahead of the other listed systems in the shown comparison.
The described training has two stages: teacher-checked memory writing followed by reinforcement learning that scores memory banks.

Chapters

Start	Chapter	Summary
00:01	Introducing Hy-Memory and the demo setup	The presenter introduces the memory-writing model, Hy Memory, and the local GPU environment used for the demonstration.
02:08	Installing and explaining Hy Memory	After installing the SDK, the presenter explains the planned test and how the system extracts, reconciles, and retrieves memories.
03:45	Connecting the model and testing memory updates	The presenter configures the local model and an embedding service, then tests whether an update changes Alex’s stored location from Seattle to Sydney.
06:50	Benchmarks and training approach	The video shows benchmark results and outlines the model’s two-stage training setup before the presenter wraps up.

Quotes

The model is the writer and this is the notebook — 00:33

We are going to connect High Memory the SDK to our local model — 02:40

The model doesn’t just pile a new note on top of the old one. — 05:48

This is a 3 billion active parameter model deciding what to keep, what to replace, and what to carry forward. — 06:19

So, the model is taught on the real job of keeping a memory bank correct over time. — 07:20

Concepts and entities

Hy-Memory-30B-A3B, Hy Memory, Tencent Hunyuan, vLLM, NVIDIA H100, SQLite, OpenClaw, Apache License 2.0, Claude Opus 5, DeepSeek V4.1

Full transcript (original language)

00:01 Memory tools for AI agents already exist. Mem zero, Mem OS and others. But they all lean on a general LLM to decide what to remember. Tencent just open-sourced a model trained for that one job with about 3 billion active parameters. Today we are going to install it, test it locally with a real-world use case. This is Fahad Mirza and I welcome you to the channel. So what exactly is this model? This is

00:33 high memory 30 billion A3B from Tencent Hunyuan. It replaces the general LLM inside memory frameworks and it decides what to keep and how to update it. Not only that, this is high memory Tencent’s own memory engine and SDK. The model is the writer and this is the notebook and we are going to install both of these. I’ll straight away take you to my terminal where I’m running this Ubuntu. I have one GPU card Nvidia

01:05 H100 with 80GB of VRAM which is consuming almost 75 gig gig at the moment due to the fact that I’m already running and serving this model on my local system with vLLM as you can see here. If you don’t know how to install vLLM, just search my channel and the reason why it is consuming so much VRAM is because of this max model length. You decrease this thing and that VRAM consumption comes down. If you’re looking to rent a GPU on very good

01:36 price, you can find the link to Mass Compute in video’s description. And here is the world’s greatest ever newsletter about AI only served to you weekly. So please subscribe to it at fahadmirza.substack.com. And if you want to help out with the channel just please become a member as that helps a lot. So now you know what is happening what exactly is running. Let me take you back to my terminal where I am running this GPU card. Now, what I’m going to do, I’m

02:08 going to install the SDK, which is primarily High Memory, and this is a single command with UV pip, and that is already done. So, you see, we have both the model installed and running, and now we have High Memory also installed. So, before I move forward, let me set the stage. So, just understand what we are going to do next. We are going to connect High Memory the SDK to our local model, right? Then we will tell it a few

02:40 facts about something, change one of those facts, and search to see whether it remembers the new version or not. That is going to show us if the model really does its job as the memory writer or not. In order to understand how exactly it all fits in, look at this diagram. This diagram shows what happens inside High Memory when we do that. Every call starts at the top, add to save something, search to get something back.

03:11 When we add a conversation, an LLM reads it and pulls out the facts worth keeping, and a reconcile step compares them with what is already stored. If a fact has changed, the old one is replaced. That LLM is the part of our model that takes over. When we search, the stored memories are recalled, and the relevant ones come back. So, when we run our test, you will know exactly which step is doing what. Let’s go back, and you know, run this thing.

03:45 The first command we need to run is this, where we are pointing High Memory’s writer at our local VLLM server. That is done. Second, I would need an embedding model because whatever memory is stored, it stores as a vector or embedding. For that, you can either use a local one or you can use an API based one. I’m just going to use the test scope API from Alibaba just to make things easier. But as I said, you can use any local model of your choice. That is also done. Now,

04:17 once that is done, I have just written a very short script to demonstrate that. Let me show you. So, look at this script. This script basically creates a high memory client in pro mode, which is the mode that uses an LLM to extract and reconcile memories. It adds a fact that Alex lives in Seattle and then adds an update that Alex moved to Sydney. Then searches for where Alex lives, so we can see whether the new fact replaced the old one or

04:47 not. So, let me go back and then run this script now. And there you go. Ran very quickly, but produced a lot of data and everything it shows which high memory it did behind one at call. So, I’ll just scroll through. Don’t worry, I’m going to explain what happened here. So, you see, um it is embedding the text, which is which means converting those memories into numerical representations or embeddings or vectors. Have the model extract facts, compare them with what is

05:18 stored, then save. The extraction picked up the Seattle to Sydney move. Um and then, you know, it is already going through through through through through through. And lot of things it has done. So, behind the scene, primarily it is using that lightweight SQLite. Okay, so let me actually trim it down. So, I have just used grep with that result just to show you what exactly is relevant here. So, you see, we tell it

05:48 Alex lives in Seattle, right? And then I’ll just scroll down to show you it has added both of the facts which we used in the script, the location and the job. And then we told it the fact too that it has uh he has moved to Sydney. And this is the moment that matters. The model doesn’t just pile a new note on top of the old one. It decides the Seattle memory is now wrong and issues a supercede with its own reason written out. The user moved from Seattle to Sydney and no longer lives in Seattle. It also updates

06:19 the job memory to carry the history along. And you see that profile at the bottom now says Alex lives in Sydney. So, this is a 3 billion active parameter model deciding what to keep, what to replace, and what to carry forward. So, this is quite impressive. And the great thing is that they have open sourced it all, which is always always welcome. Apache 2 license. How good is that? You can even integrate it with Open Claw or any other coding agent,

06:50 too. I think Open Claw integration at a scale is already there in their GitHub repo. Before I wrap up, let me show you quickly one more thing. They have shared few of the benchmarks. So, if you look at this diagram, they tested four memory writers on the same setup. And the only thing that changed was which model wrote the memories. And you can see High Memory, this one, scored 55.49 overall, which is ahead of Claude Opus 5, DeepSeek V4.1, and the

07:20 size mashed Qwen net 50.37. Really really cool. Um let me very quickly show you the training setup. So, it has two stages. It’s in stage one, the student model writes memories and a frozen H High 3 teacher checks every step it takes. In stage two, reinforcement learning scores the memory bank after each chunk using a frozen answer model and a judge. And the best bank carries forward. So, the model is taught on the real job of keeping a

07:52 memory bank correct over time. So, look. Um really impressive, very very liberal and generous license. I think I will be using this tool in my production environment with my client. So, feel free to check it out. Let me know your thoughts. And again, please become a member and do subscribe to that newsletter. Thank you for all the support.