RecurrentGPT is a model designed for the interactive generation of text of any length. It replaces vectorized elements in Long Short-Term Memory (LSTM) networks with natural language (i.e., text paragraphs) and simulates a recursive mechanism through prompt engineering. At each time step, RecurrentGPT receives a text paragraph and a brief plan for the next paragraph, both generated in the previous time step. It also maintains short-term memory to summarize key information from recent time steps and updates it at each time step. RecurrentGPT combines all inputs into a prompt, requesting the foundational language model to generate new paragraphs, provide brief plans for the next sections, and update long-term and short-term memory.