AnswersPublished17 sources
What Tencent Open-Sourced in Hy Memory 2.0
Hy Memory 2.0 is a modular AI agent memory stack: a framework and Python SDK, the Hy Memory 30B A3B model, and MemHub. The framework organizes chat, working and code memory; developers can connect it to agents through the SDK.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Hy Memory 2.0 is a modular AI agent memory stack: a framework and Python SDK, the Hy Memory 30B A3B model, and MemHub.
- The framework organizes chat, working and code memory; developers can connect it to agents through the SDK.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Tencent Hunyuan open-source in its October 17, 2026 release of Hy Memory 2.0, and how do its independently usable or combinable fra. Article summary: Tencent Hunyuan open-sourced three layers of Hy Memory 2.0: the Hy-Memory agent-memory framework and Python SDK, the dedicated Hy-Memory-30B-A3B model, and the MemHub local memory application, usable separately or togeth. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Tencent Hunyuan’s Hy Memory 2.0 is an open-source, three-part memory stack for AI agents: a framework and Python SDK, a specialized memory model, and MemHub, a local-first app. Each layer can be used on its own or combined. The wider announcement was reported on October 16, 2026; the project changelog dates the hy-memory==2.0.2 package release to October 17. 16
21
30
Three layers, three roles
1. Hy-Memory framework and Python SDK
The framework manages the storage, updating and retrieval of three kinds of memory: chat, working and code. Developers can use the SDK to integrate it with AI agents. The project changelog lists hy-memory==2.0.2 as published on October 17, with Apache-2.0 package metadata. 16
21
30
The categories extend beyond conversation history: working memory relates to tasks, while code memory addresses development work. The available source material describes these as supported memory types, but does not establish every detailed behavior claimed in some reports—such as exactly how the system records task decisions or handles memory before work resumes. 14
16
2. Hy-Memory-30B-A3B model
Tencent also released Hy-Memory-30B-A3B, a mixture-of-experts model described as having about 3 billion active parameters and trained for memory extraction and conflict handling. Its weights were published on Hugging Face with a technical report. 1
14
This model is a distinct component of the stack, rather than the framework itself. Tencent’s product page reports an average score of 55.49 for the model in its evaluation and compares it with Claude Opus 5 at 54.62. That is a company-reported result; it should not be confused with the framework’s separate memory-task benchmark scores. 16
3. MemHub local app
MemHub offers command-line and Mac desktop options. The repository describes it as local-first software that brings conversation history and coding traces from user-authorized AI applications into a managed runtime and local data directory. Documented examples include Claude Code, Cursor, Codex, OpenCode and Qwen Code. 4
14
That setup is intended to make memory available across supported tools while keeping the memory store local. The supplied sources establish the local-first design and listed integrations, but do not verify every claim about the number of connected tools or ongoing synchronization of all past and new sessions. 4
Reported internal use and benchmark results
Tencent said Hy Memory was already being used in internal finance, news, knowledge-management, office-agent, social-entertainment and gaming work. This is reported company adoption, not an independent assessment of its impact. 21
Tencent also says Hy Memory ranked first among the open-source frameworks it compared across four benchmarks. Reported results include: 16
35
- MemoryArena: 55.31% average performance on travel-planning and combined-shopping tasks; a report says this was 10.2 percentage points ahead of the next-ranked system.
35
- CodeMemoryBench: 64.2% code question-answering accuracy and a 35.6% code-task resolution rate.
- SWE-ContextBench: 28.3% code-task resolution.
- LoCoMo: 72.6% chat-memory accuracy. The same report gives 1,159 tokens per question for the comparison system and 703 for Hy Memory, a reported reduction of 39.3%.
These are reported benchmark comparisons, not independent validation. The figures describe different tasks and measures, so they should not be read as a single overall score or proof of performance in every deployment. 16
35
What the release means for agent builders
Hy Memory 2.0 packages agent memory as components rather than a single all-or-nothing product. Developers can start with the framework and SDK, use the dedicated model as a separate memory-writing layer, or add MemHub for local cross-tool memory management. Tencent’s reported internal use and benchmark results show how the company positions the system; teams evaluating it should still test their own workloads and verify which MemHub integrations and memory behaviors meet their requirements. 4
14
16
21
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Tencent Open-Sourced in Hy Memory 2.0"?
Hy Memory 2.0 is a modular AI agent memory stack: a framework and Python SDK, the Hy Memory 30B A3B model, and MemHub.
What are the key points to validate first?
Hy Memory 2.0 is a modular AI agent memory stack: a framework and Python SDK, the Hy Memory 30B A3B model, and MemHub. The framework organizes chat, working and code memory; developers can connect it to agents through the SDK.