AirJelly launches a desktop proactive AI assistant, capturing user intent with the Enter key as an anchor point

ME News update: On April 22 (UTC+8), according to Dongcha Beating monitoring, AirJelly, a desktop AI assistant released by persistent low-entropy (Low Entropy AI) founded by former ByteDance product manager Bert, has secured funding from Siyuan Capital. Bert previously led ByteDance’s context engineering product MineContext, then left with his original team to start a business. The team consists of 11 members, and all of them work offline at their office in Beijing.

AirJelly does not do full-screen recording. Instead, it uses the Enter key as an anchor: every time the user presses Enter, it captures the screen, and through Accessibility permissions it obtains the current app, the type of input field, and the context. These are modeled as Events, and then summarized into Tasks by AI. Compared with MineContext’s previous full-screen approach, which took about 1,500 screenshots per day, the Enter mechanism compresses the screenshot volume to about 300, bringing costs down to one-fifth, while also reducing misjudgments caused by irrelevant screenshots.

The memory system is structured in two layers: static information is modeled as Entities (people, projects, etc.), and dynamic information is modeled as Tasks. Each Task includes a title, summary, progress, next steps, and related Events. During recall, it combines vector retrieval, keyword search, and time-decay weighting, with all data stored locally. On this basis, AirJelly determines whether a Task update reaches the push threshold, and proactively recommends the next action to users—a process the team calls Proactive Trigger.

At the underlying agent execution layer, the capability is integrated with OpenClaw’s Pi framework, combined with MineContext’s screen understanding ability. The product currently supports macOS. Windows and Linux versions are under development and can be downloaded for free at airjelly.ai. The team’s next plan is to launch the “Next Enter Prediction” feature, which predicts what users will enter next when pressing Enter based on historical behavioral trajectories, and to introduce a team version that supports multi-person collaboration.

(Source: BlockBeats)