Tencent open-sources Hy-Memory 2.0 to give AI agents shared memory

Hy-Memory 2.0 is an open-source memory architecture from Tencent designed to solve AI context fragmentation. It ships as a full three-tier stack: a dedicated memory framework, an accompanying desktop application called MemHub, and a 30B specialized model.

Basically, it lets you bounce between Cursor, Codex, WorkBuddy, and Hermes without re-explaining your project specs every single time.

What it does and how it works

The standalone app, MemHub, connects up to 10 different AI tools. You can authorize your existing apps, ingest past sessions via a local SDK, and build a unified local memory pool. New interactions from any connected client sync in real time, making past decisions instantly visible to whichever agent you open next.

Under the hood, the framework divides memory into three specific modules instead of treating all text the same. Chat Memory handles user profile details and resolves conflicting statements across sessions. Working Memory separates task milestones from reusable recipes, explicitly tracking which solution worked under what constraints so agents only reuse proven steps when conditions match. Coding Memory captures git diffs, historical context, and file evolution, giving coding agents the exact rationale behind past refactors.

Dedicated 30B weights and benchmarks

Alongside the engine, Tencent released Hy-Memory-30B-A3B, a model fine-tuned specifically to filter, structure, and update long-term memory banks. It can plug into Hy-Memory or third-party systems like Mem0 and MemOS.

Across six long-term memory evaluations including LoCoMo and LongMemEval, the model averaged 55.49 points across four distinct memory harnesses, edging out Opus5 at 54.62 and DeepSeek V4.1 Flash at 53.64.

On the framework side, Hy-Memory topped tests like MemoryArena with an average score of 55.31% on travel and shopping benchmarks, leading runner-up EverOS by 10.2 percentage points while keeping memory bank token consumption low.

The MemHub app and integration code are hosted on GitHub, with the 30B model weights available directly on Hugging Face.

Why it matters

For developers constantly jumping between specialized coding tools and chat assistants, Hy-Memory tackles the daily friction of context switching. Isolating code diff chains and reusable task logic is much smarter than dumping raw chat logs back into context windows.

Decoupling memory from proprietary platforms turns your personal context into a portable asset. While local synchronization and real-time retrieval accuracy across multiple third-party apps will determine how well it holds up in messy real-world projects, this release gives open-source AI workflows a genuinely practical backbone.