Tencent's Hunyuan team has put the whole stack behind its agent memory system into the open. Hy-Memory 2.0, presented on the project's official site with a release date of October 16, 2026, comes in three parts that work alone or together: a model trained to write long-term memories, an engine and Python SDK that store and retrieve them, and a desktop and command-line app called MemHub. The code lives in a new Tencent-Hunyuan/Hy-Memory repository on GitHub, created on October 15, and Tencent says it is released under the Apache 2.0 licence.
The model, Hy-Memory-30B-A3B, is a mixture-of-experts network with 30 billion parameters, of which roughly 3 billion are active at a time. Its weights appeared on Hugging Face on October 15, together with a technical report on arXiv. It is not built to chat with users. Its job is to replace the general-purpose LLM that memory frameworks call when they decide what to keep, how precisely to record it and how new evidence should revise what is already stored. Tencent trained it across three frameworks, Mem0, MemOS and its own Hy-Memory, first with on-policy distillation and then with a reinforcement-learning stage that rewards individual memory writes according to whether later questions can be answered from them.
According to the report, the model ranks first on all four memory harnesses Tencent tested, including EverMemOS, which was kept out of training, with a macro-average of 55.49% against 54.62% for Claude Opus 5. Those results come from fixed 20% subsamples of a benchmark set that covers LoCoMo, LongMemEval, PersonaMem v2, two MemoryArena tasks and CMB, a coding-memory benchmark built by the same team. They are Tencent's own numbers, and no outside replication has been published.
The engine layer ships on PyPI as hy-memory, at version 2.0.4 when this story was checked. It treats memory as three separate scenarios. For chat, personal information is arranged in a hierarchy that runs from raw dialogue through atomic facts, summaries and schemas up to future intentions, a design described in a paper accepted at EMNLP 2026. For task-oriented agents, a working-memory mode keeps one layer of facts and another of reusable experience. For coding, a chain of version diffs lets an agent recall not only what the code looks like today but why it changed. On its project page Tencent says the engine came out ahead of Mem0, MemOS, EverOS, OpenViking, RepoMemory and a simple RAG baseline in its chat, working and coding evaluations.
MemHub is the piece meant for people who already juggle several AI assistants. It imports earlier conversations and coding traces from applications the user authorises, such as Claude Code, Cursor, Codex, OpenCode and Qwen Code, keeps them in one local data directory and, through plugins, feeds the relevant memories back into whichever tool is active. The project page counts ten supported tools. MemHub v1.0.4, published on October 18 with a signed macOS app and verified macOS and Linux runtimes, followed the first v1.0.0 tag two days earlier.
Hy-Memory is not a new name: version 1.0 of the framework appeared in May 2026. The difference with 2.0 is that Tencent is releasing a trained memory model, the framework and an end-user app in one go, at a time when persistent memory has become a recurring weakness of coding agents that forget their context between sessions. Three days after launch the repository had about 900 stars on GitHub, and the model card listed close to 1,200 downloads.
Several questions remain. Every performance figure is self-reported, and the comparison with Claude Opus 5 relies on subsampled benchmarks, one of which Tencent designed itself and has only submitted for review at ICLR 2027. In the framework comparison, Tencent used DeepSeek V4.1 Flash as the memory model and GPT-5.6-sol to answer and judge, so results may shift with other model choices. How well MemHub's cross-tool recall holds up also depends on how each third-party application stores its history, and no independent assessment of its reliability or privacy controls has appeared yet.