Portable agent memory from Tencent: the Hy4 preview squeezes 1.5 terabytes into 214 gigabytes
In the Hunyuan Hy4 preview, agent memory travels across Claude Code, DeepSeek Harness and WorkBuddy; a 1.5 TB memory shrank to 214 GB with almost no benchmark loss, running on a single 4090 laptop.
Tencent's AI team announced the Hunyuan Hy4 preview with an unusual emphasis: memory, over model. In the team's words, most teams ship an open model and call it done; what Hy4 brings is agent memory that travels across Claude Code, DeepSeek Harness and WorkBuddy. Instead of being locked into a single vendor's database, the memory moves with the agent.
The numbers land too: 1.5 terabytes of agent memory compressed to 214 gigabytes with almost no change on benchmarks, and the package runs on a single 4090 laptop or a box of four A4000s. On code, the Hy4 preview climbed to fifth on Code Arena's WebDev ranking; Hy3 sat at 31st two months ago. Virtual environments get attention as well: an agent can be paused on one machine and woken on another with its state intact. All of it open source.
The context is familiar: we covered Tencent's ContextPilot yesterday, which teaches context management to the model; Hy4 is the memory front of the same war. It also makes the lock-in debate of the agent era concrete: an agent that carries its memory in its pocket can move to any harness it likes, and what keeps the customer has to be the service itself, no longer the database.