The Companion Cube

## Reto This project investigates AI memory, treating it not as a neutral mechanism but as a point of vulnerability and user intervention. By researching how large language models (specifically OpenAI’s ChatGPT) filter, forget and shape information, I developed the Companion Cube: a device that translates invisible AI memory states into expressive visual warnings users can intuitively read. The aim is not to fix AI, but to change our relationship with it: to expose its limits, interrupt the illusion of understanding, and give users more literacy, agency, and control. ## Proceso The project began with Context Void, an investigation into how large language models retain, filter, and lose information over long conversations. Through literature review and empirical testing of ChatGPT, I identified recurring patterns of context degradation and translated these findings into measurable indicators. The project evolved from data visualisations to a Chrome extension, before becoming the Companion Cube: a physical Arduino-based device that communicates with ChatGPT through a desktop application. Each design decision focused on making invisible AI memory states visible, intuitive, and actionable while encouraging users to critically engage with conversational AI. ## Resultado The outcome is the Companion Cube, a functional companion device that translates invisible AI memory states into expressive visual warnings users can intuitively read. Rather than improving AI itself, it improves users' literacy by revealing the limits of AI memory and interrupting the illusion of understanding. The project demonstrates how critical design can make complex AI systems more transparent while encouraging better prompting practices and greater user agency. Future work will refine the trigger system, simplify the hardware, and explore broader applications for communicating AI reliability across different interfaces.