Tencent describes Hy-Memory 2.0 as a cross-tool agent-memory layer released to the public. The package includes a memory-writing model, an SDK and framework, and what the company calls a local-first application named MemHub. That changes the immediate question for developers from whether an agent can retain a conversation to whether a task can carry the right context across chats, productivity work, and coding tools without turning personal history into an unreadable data cache.
The release comes from Tencent’s Hunyuan model team, which describes the project as an open-source release on its Tencent Hunyuan launch page. The public GitHub repository, an installable hy-memory 2.0.1 package on PyPI, and an ungated Hugging Face model page make the release more tangible than a product announcement. Developers can inspect code, obtain package artifacts, and evaluate public model weights rather than wait for a hosted feature to arrive.
Hy-Memory 2.0 Turns Agent Memory Into a Usable Stack
The release is organized as three layers that can be used separately or together. Hy-Memory-30B-A3B is the memory-writing model. The Hy-Memory engine and SDK provide the software layer around it. Project documentation describes MemHub as a desktop and command-line application intended to collect memory from user-authorized AI applications. Tencent says the stack addresses three forms of memory: chat, working, and coding.
That division matters because retaining a chat transcript is only one form of continuity. The project documentation says working and coding memory use tool-call traces, while coding memory can preserve a chain of version changes through local Git differences or parseable replacement actions. In a resumed task, those records could give an agent more than a final answer to work from: they could supply the choices and revisions that led there. The documentation describes the design; it does not establish that a given agent will preserve the right information in ordinary use.
The public materials give developers an evaluation starting point. The Hugging Face repository lists 51 Safetensors shards, roughly 30.06 billion total parameters, and about 3 billion active parameters. PyPI offers the non-yanked 2.0.1 SDK for Python 3.9 or later.
Tencent previously explored Hy-Memory, and the new release is expressly presented as a broader step beyond the earlier chat-focused, non-open-source 1.0. Readers who saw Tencent Hunyuan’s earlier Hy-Memory plugin for agent collaboration should see 2.0 as an expansion in what is available to inspect: it adds a public model, framework, and application rather than a single agent-memory feature. The associated Hy-Memory-30B-A3B technical report is a public Tencent Hunyuan preprint rather than a peer-reviewed result.
MemHub Makes Permission and Deletion Part of the Product Question
MemHub is where the release reaches beyond model evaluation. Tencent’s project documentation describes it as local-first and says it consolidates historical and new data from user-authorized AI applications into a personal memory layer. Tencent says it can share that memory across 10 or more tools. For someone moving among a chat assistant, a productivity tool, and a coding environment, the promised benefit is continuity without manually restating the task at every handoff.
That promise also gives the reader a concrete set of questions before attaching the software to sensitive work. According to the project documentation, MemHub’s default SQLite data live in ~/.one-memory. The same documentation says the product does not collect screens, accessibility trees, OCR, keyboard or mouse activity, clipboard contents, or background audio. These are Tencent/project design statements, not an independent examination of the software. The useful test is therefore specific: identify which histories are imported, which tool plugin receives permission, what can later read the stored memory, and how the stored record is removed.
Deletion is especially material for a product built to preserve task context. The documentation says users must explicitly install the complete runtime and that only the delete-data command removes stored data. That is an operational detail, not fine print. A developer evaluating cross-tool memory needs to understand whether an experiment leaves traces behind, where those traces reside, and whether cleanup behaves as described. The released application is supported on Linux and macOS builds; the documentation explicitly says Windows is unsupported.
This is also why the story is about more than a model leaderboard. A cross-tool memory layer connects separate pieces of work and, if it works as designed, makes a long-running task easier to resume. It also makes the boundaries around imported histories a product decision. The question of what a tool retains may be familiar to readers of EastFrontier’s agent-access coverage. Hy-Memory 2.0 gives developers a released stack through which to test Tencent’s stated local-first design rather than treat memory as an invisible feature of one assistant.
Tencent’s Memory Scores Need a Reproducible Field Test
Tencent reports strong results, but the useful way to read them is as a test plan. On its release page, Tencent says the framework ranked first among open-source frameworks on four named benchmarks spanning its three memory settings. Its technical report and model card report a 55.49 macro-average for Hy-Memory-30B-A3B, compared with 54.62 for Claude Opus 5 across four harnesses. The claim is narrowly defined by the experiment: the memory writer changes while prompts, retrieval configuration, answer model, judge, and other framework components are held fixed.
The numerical qualification is central. Tencent’s report says each benchmark in that four-harness matrix used a fixed 20% subsample, each harness averaged six metrics equally, and the overall score averaged the four harnesses equally. The paper itself points to unfinished validation work, including a random human review, checks for duplication or leakage, calibration by scenario and failure type, and sensitivity testing for the judge. That leaves independent reproduction as the real measure of whether the public release carries its reported advantage into a developer’s workload.
A meaningful comparison would run the released 2.0.1 SDK and public Hy-Memory-30B-A3B weights against at least one alternative memory framework on a disclosed workload. The workload should include long, multi-session chat facts that later conflict; a task resumed after a tool change; and a coding task containing a reverted edit. Holding the answer model, embeddings, retrieval depth, prompts, hardware and cost budget, and test set steady would expose the memory layer rather than the surrounding system. Full-task results, memory-bank size, and error categories would matter more than a single average.
MemHub needs its own practical test on the macOS or Linux builds actually released. An evaluator can inspect opt-in imports, separation of sources, local data location, tool-plugin permissions, and deletion behavior while resuming a task across tools. That paired test would answer the two questions Tencent has put before developers: whether the memory-writing stack retains useful context, and whether its cross-tool continuity remains understandable and controllable when real work, rather than a benchmark harness, is on the line.