Intensive care unit (ICU) patients often develop new health-related problems in their long-term recovery. Health care professionals keeping a diary of a patient's stay is a proven strategy to tackle this but faces several adoption barriers, such as lack of time and difficulty in knowing what to write. Large language models (LLMs), with their ability to generate human-like text and adaptability, could solve these challenges. However, realizing this vision involves addressing several socio-technical and practical research challenges. This paper discusses these challenges and proposes future research directions to utilize the potential of LLMs in ICU diary writing, ultimately improving the long-term recovery outcomes for ICU patients.