Hy-Memory-30B-A3B

Hy-Memory-30B-A3B

Writes an assistant's long-term memory, choosing what to keep and how to update it inside Mem0, MemOS or Hy-Memory, with about 3B active parameters. For developers building AI agents with memory.

Made by tencent on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Hy-Memory-30B-A3B is Tencent Hunyuan's model for the writing side of agent memory, with about 3B active parameters out of 30.1B. In Mem0, MemOS and Hy-Memory it takes the place of the general language model. It picks which facts to keep, how exactly to record them, and how fresh evidence should change what is already stored.

Who it suits: It suits developers who build assistants or coding agents on a memory framework and want a dedicated model to decide what goes into the memory bank.

Industry | Who uses it | How they use it | Example
Agent development | AI engineers | Swap it in as the memory writer in a memory framework | Replacing the general model inside Mem0
Chat products | Product developers | Keep durable facts about users across conversations | Updating a stored preference after a new message
Coding tools | Tool builders | Keep working notes about a codebase between sessions | Using the Hy-Memory coding scenario
Research | Memory researchers | Compare memory writers under identical conditions | Reading the card's cross-framework table

If you only need a chat model, or you do not use a memory framework, you can skip it.

Where to use it:

• Mem0, MemOS and Hy-Memory, which cover chat, working and coding scenarios
• EverMemOS, which the card says it carries over to without extra training
• Python with transformers, on Windows, macOS or Linux

Use case | Who | What they get back
Chat memory | Assistant builders | Stored facts that persist between conversations
Working memory | Agent builders | Notes on the task in progress
Coding memory | Tool builders | Stored knowledge about a repository

How to set it up:

• Python and the Hugging Face hub package
• The transformers library, which the model page lists
• A memory framework such as Mem0, MemOS or Hy-Memory
• About 60.1 GB for the BF16 weights, or about 15 GB at 4-bit by the card's estimate

1. Install the Hugging Face hub package.
2. Download the model files with the command below.
3. Load the model with transformers.
4. In your memory framework, use this model wherever it asks for the language model that writes memories.

pip install -U huggingface_hub
hf download tencent/Hy-Memory-30B-A3B

How to use it: You do not chat with it directly. Your memory framework sends it conversation or work content, and the model decides what to write into the memory bank and how to revise old entries.

Tool | What it does
Mem0 | A memory framework that this model can serve as the writer in
MemOS | Another memory framework with the same swap
Hy-Memory | Tencent's memory framework, covering chat, working and coding scenarios
EverMemOS | A framework held out from training, which the card says the model still handles

Example: You run an assistant on Mem0 and swap this model in as the memory writer. After a conversation it stores the facts that matter, and when a user later changes a detail, it revises the stored entry.

For newcomers and experts: Assistants forget everything between chats unless something writes notes into a memory store. This model is that note-writer. It does not answer users itself. A good first try is to swap it into one memory framework you already run and compare the memories it saves.

Under the hood, it was trained in two stages. Memory On-Policy Distillation came first, then Attributed Memory Reinforcement Learning. The card tests it by changing only the memory writer, while the answer model, the judge and the retrieval setup stay fixed. Each framework score averages six metrics, among them LoCoMo and LongMemEval.

Conversation or work log comes in
Writer decides what to keep
Memory bank is updated
Later questions read from the bank

Good to know:

Title: Average score across four frameworks, higher is better
Hy-Memory-30B-A3B: 55.49 points
Claude Opus 5: 54.62 points
DeepSeek V4.1 Flash: 53.64 points
Hy3: 51.63 points
Qwen3.6-35B-A3B: 50.37 points
Hy3-A3B Base: 47.66 points
Qwen3-30B-A3B: 44.37 points

Title: Score by memory framework, higher is better
Memory writer | Mem0 | Hy-Memory | MemOS | EverMemOS
Hy-Memory-30B-A3B | 54.02 | 60.55 | 50.51 | 56.89
Claude Opus 5 | 53.43 | 60.53 | 50.50 | 54.04
DeepSeek V4.1 Flash | 49.73 | 60.20 | 49.17 | 55.47
Hy3 | 45.99 | 58.06 | 50.19 | 52.30
Qwen3.6-35B-A3B | 47.51 | 55.87 | 47.36 | 50.75
Hy3-A3B Base | 40.20 | 50.39 | 49.05 | 50.98
Qwen3-30B-A3B | 39.81 | 45.67 | 45.44 | 46.57

• The scores come from fixed 20% subsamples of the benchmarks, so they are a sample and not the full sets.
• Claude Opus 5 scores higher on parts of the card's Mem0 breakdown, such as LoCoMo at 69.47 against 59.54.
• EverMemOS was held out from training, and the card still reports a result for it.
• Other people have published 24 quantized and 2 fine-tuned versions of it.
• License: Apache 2.0
• Cost to run: none named, though the BF16 weights take about 60.1 GB
• Data it reaches: the conversation content your framework sends it, and the memory bank it updates
• Not for: none named, though it is built as a memory writer and not as a general chat model

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 30.1B
• Weights: about 60.1 GB in BF16; about 15 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,755
• Likes: 325
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 24
• Fine-tunes: 2
• Tags: transformers, safetensors, hy_v3, text-generation, conversational

Links

• Model: https://huggingface.co/tencent/Hy-Memory-30B-A3B
• Files: https://huggingface.co/tencent/Hy-Memory-30B-A3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/Hy-Memory-30B-A3B

Then follow the usage in its model card. Nothing in this kit runs by itself.