> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Tencent open-sources Hy Memory 2.0 and a model that writes agent memory
- URL: https://globalfeed.ai/en/tencent-open-sources-hy-memory-2-0-and-a-model-that-writes-agent-memory/
- Published: 2026-10-19T00:06:58.000Z
- Updated: 2026-10-19T00:06:58.000Z
- Description: Tencent Hunyuan has open-sourced Hy Memory 2.0, a long-term memory stack for AI agents. It includes a Python SDK, the MemHub app that gathers coding-tool histories in one local place, and an Apache 2.0 model that decides what goes into memory.
- Author: GlobalFeed Editor
- Tags: Tencent, Hunyuan, Hy Memory, agent memory, open model, x-TencentHunyuan, dil-en, elle

Tencent Hunyuan released [Hy Memory 2.0](https://github.com/Tencent-Hunyuan/Hy-Memory?ref=globalfeed.ai) as open source in October 2026\. Hy Memory is a memory layer that lets AI agents keep track of conversations, tasks and code changes across sessions. The release has three parts: the Python SDK `hy-memory`, the MemHub desktop and command-line app, and [Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B?ref=globalfeed.ai), a model trained specifically to write memory. The model, which you can run on your own servers, is on Hugging Face under the Apache 2.0 license.

## What Hy Memory does

An agent that runs for a long time has to decide what to remember, what to update and what to drop. Hy Memory handles this with separate designs for three scenarios: in chat it keeps personal facts in a layered long-term memory, in work tasks it stores facts and reusable experience, and in coding it tracks how code changes version by version in a diff chain. According to the repository, the SDK only needs a language model and an embedding endpoint; any OpenAI-compatible service or a local model can be used.

## MemHub: shared memory for coding tools

MemHub collects conversation history and coding traces from AI apps the user authorizes into a single local folder. The repository lists Claude Code, Cursor, Codex, OpenCode and Qwen Code as examples, so something learned in one tool can be recalled in another. Data stays on the user's own machine; a signed macOS desktop app and command-line runtimes for macOS and Linux are available.

## The memory-writing model

Hy-Memory-30B-A3B is a mixture-of-experts model with 30 billion total parameters and about 3 billion active per step. Its job is to decide what gets written to memory and how new evidence updates old entries. The model card says it can directly replace the general-purpose LLM inside Mem0, MemOS and Hy Memory. In the card's table its cross-framework average is 55.49, against 54.62 for Claude Opus 5 and 53.64 for DeepSeek V4.1 Flash. It runs with vLLM, SGLang and Transformers, and quantized builds for llama.cpp, Ollama and LM Studio are listed. The training method is described in the [technical report](https://arxiv.org/abs/2610.22196?ref=globalfeed.ai).

The scores are values shared by Tencent; the tests used fixed 20 percent subsamples of the datasets, and the card notes that some framework results were not collected the same way for every model. The card gives no detail on language support, including Turkish. Trying it in your own agent workflow and following independent researchers' comparisons would be useful.