Savindu Thathsara is developing an offline AI operating system for 8GB laptops, focusing on personalized learning. The AI assistant learns user habits locally, without cloud connectivity or large API models. Its core loop involves noticing actions, storing short records, and transforming repeated actions into habits. The system offers an easy version of learned actions or performs safe actions when requests are vague. Corrections from the user update these records. Accidental events are not immediately established as habits; repetition is key. The memory fills in details for existing skills, such as playing a specific song or opening a designated photo folder, but does not invent new programs or shell commands.The assistant learns from on-screen activity by reading the Linux accessibility tree, capturing application names, window titles, and visible controls. It samples this information when the front window changes or during user activity, but avoids recording video, screenshots, or sensitive data like passwords. The data stored includes spoken or typed commands, app and window details, file paths, system settings like volume, and skill execution status. Importantly, it excludes keystrokes, document bodies, chat content, and any cloud copies of the data. Users can view and delete learned habits, which are presented as clear sentences, and can pause learning, including screen reading. Risky actions still require user confirmation, even if they match a known pattern.The system is designed to remain small and efficient. It avoids loading large vision models; instead, it queries the existing accessibility tree. Habits are based on counts and recent examples, not additional large neural networks. The entire database resides on the laptop, ensuring privacy and local operation. The current focus is on implementing memory with typed commands, with voice integration planned for later. Thathsara, based in London, is openly documenting this project, including how local LLMs, speech-to-text, and wake word detection will operate within the 8GB RAM constraint alongside a web browser.