Tencent Open-Sources Hy Memory 2.0 for AI Agent Recall

AI agents keep forgetting things. Every coding assistant, chat agent, and office helper stores what it learns inside its own app, and when the work moves to a different tool, the context has to be explained all over again. On October 16, Tencent's Hunyuan team released Hy Memory 2.0, an open-source memory framework designed to give AI agents a shared, persistent memory that survives across applications and sessions.

Hy Memory 2.0 is not a single library. According to the announcement reported by WallstreetCN, the release covers three layers at once: a dedicated memory model, the memory framework itself, and MemHub, a local application that pulls conversation histories from multiple AI tools into one shared system. Each layer can be used on its own or combined into a full memory stack. According to the project's GitHub page, the public code and model weights are released under the Apache 2.0 license.

What the three layers do

The framework inside Hy Memory 2.0 splits agent memory into distinct pipelines instead of treating everything as one undifferentiated chat log. According to coverage of the launch, the design separates three "memory scenes": hierarchical long-term memory for chat, a two-layer system for working tasks that stores both facts and reusable lessons, and version-evolution tracking for code changes that records why a change was made, not just what changed. Tencent presents this separation of working memory from code memory as a first for the field, which is the company's own framing of the advance.

Sitting underneath the framework is a specialized memory model, a 30-billion-parameter mixture-of-experts model released on Hugging Face alongside a technical report. According to an analysis of the release by the independent outlet alextech, the model is trained in two stages: an initial phase that teaches each memory framework's write protocol, followed by reinforcement learning that rewards memory writes which help answer future questions. The company reports that the model averages 55.49% across four memory frameworks it tested, ahead of Claude Opus 5 at 54.62% and DeepSeek V4.1 Flash at 53.64%. Those figures are Tencent's own and have not been independently verified, but the more practical claim is about cost: with roughly 3 billion activated parameters, the memory-writing role can be handed to a much smaller model instead of the expensive frontier model doing the reasoning.

The third layer, MemHub, is the one aimed directly at the daily friction of using AI tools. According to the project's documentation, MemHub is a local-first desktop and command-line application that unifies conversation history and coding traces from user-authorized AI applications — including Claude Code, Cursor, Codex, OpenCode, and Qwen Code — into a single managed local data directory. Coverage of the launch puts the number of supported tools at 10, with integrations requiring the user's authorization.

The problem it is trying to solve

The motivation is straightforward and familiar to anyone who supervises AI coding agents. A developer might explain a project's architecture to one coding agent, set preferences in a different assistant, and document a decision in a third tool — then repeat all three when switching between them. As one recent report on the launch put it, context remains trapped inside individual applications, and Hy Memory 2.0 aims to make relevant facts, task progress, and the reasoning behind code changes portable instead.

This is the same coordination gap that other October launches have been chasing from different directions. BAND for Developers arrived earlier this month to help developers orchestrate fleets of coding agents across tools and machines; a fleet like that is exactly the kind of setup where shared memory matters, since agents handing work to each other need a common record of what was tried and decided. Meanwhile, MAI-Gym offers an open platform for training and evaluating GUI agents. The infrastructure around agents — orchestration, evaluation, and now memory — is being built out in parallel.

According to the WallstreetCN report, Hy Memory is already in use across multiple Tencent businesses, including finance, news, knowledge management, office agents, social entertainment, and gaming. That internal deployment is a more meaningful signal than the benchmarks: a memory layer that survives production use inside a large company has cleared a bar that demo videos cannot.

Early evidence, with caveats

One independent developer who works on memory systems tested Hy Memory 2.0 against the LongMemEval benchmark shortly after release, focusing on the two failures that matter most in practice: answering with an outdated version of a fact that has since changed, and forgetting who said what — including what the assistant itself said. The tester's headline finding was that the framework's performance depends heavily on the language model doing the memory writing, which is worth remembering when evaluating the company's benchmark claims. A memory framework is only as disciplined as the model filling it.

There are open questions that only broader adoption can answer. The reported benchmark figures are company-supplied. MemHub's value depends on how smoothly its integrations actually authorize across the supported tools, and on whether developers accept a local hub as the single place their agents' histories live. The technical report is on arXiv, the code is on GitHub, and the model is on Hugging Face — so the inspection part is already done. Whether Hy Memory 2.0 becomes shared infrastructure or another research artifact will be decided by the people who try wiring it into their agent stacks. The alextech breakdown of the release provides a deeper look at the training approach for those who want to judge it themselves.

One more data point from the launch coverage: the framework's design assumes that bigger context windows alone will not fix agent forgetfulness. As the Lapaas Voice report on the release noted, the pitch behind Hy Memory 2.0 is that a purpose-built memory layer, with a model trained specifically to write memories, beats just stuffing more tokens into the prompt. That bet — structure over scale — is the one to watch as agent memory systems compete to become the standard layer under every agent framework.