Agents always "lose memory" on long tasks? Domestic open source breaks the deadlock, Jev-like models take over Agent scheduling.

What happened: APUS open sourced Reflex Agent under the MIT license, offering desktop, Web, and SDK versions. It uses a self developed Reflex decision model to drive the Harness control loop, with a three layer memory system for unlimited context. Key numbers: APUS founded 2014, valued about 12 billion yuan in 2017. Why it matters: The approach separates scheduling decisions from the large model and preserves task details through lossless memory, offering a different technical path for long task agents. Large scale deployment still needs community validation.

APUS open sourced Reflex Agent under the MIT license, offering desktop, Web, and SDK versions. It uses a self developed Reflex decision model to drive the Harness control loop, with a three layer memory system for unlimited context.

The approach separates scheduling decisions from the large model and preserves task details through lossless memory, offering a different technical path for long task agents. Large scale deployment still needs community validation.