Bai Te, a former product manager at ByteDance, has launched a desktop AI assistant called AirJelly through his startup Low Entropy AI. The company has already secured funding from 5Y Capital (formerly Morningside Venture Capital). Bai previously led the development of MineContext, a context engineering product at ByteDance, before leaving with his original team to start this venture. The team of 11 people all work together in an office in Beijing.
https://twitter.com/airjellyAI/status/2046621478662402498
AirJelly doesn't do full screen recording. Instead, it anchors on the Enter key: every time you press Enter, it captures a screenshot. Through accessibility permissions, it also grabs the current app, input field type, and context. It models this information as an Event, which the AI then summarizes into a Task. Compared to MineContext's full-capture approach that took about 1,500 screenshots per day, the Enter mechanism cuts that down to roughly 300 screenshots, reducing costs to one-fifth and cutting down on false positives from irrelevant screenshots.
The memory system has 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 step, and associated Events. When recalling, it combines vector search, keyword search, and time-decay weighting, with all data stored locally. On top of that, AirJelly decides whether a Task update meets a push threshold and proactively suggests the next action to the user — the team calls this "Proactive" triggering.
https://www.youtube.com/watch?v=H1DSG__lay4
Under the hood, the agent execution capability taps into OpenClaw's Pi framework, combined with MineContext's screen understanding abilities. The product currently supports macOS, with Windows and Linux versions in development. It's available for free download at airjelly.ai. Next up, the team plans to launch a "Next Enter Prediction" feature that predicts what you'll type next based on historical behavior, as well as a team edition for multi-user collaboration.