Multi-Agent Framework for Interactive Storytelling using Large Language Models

DreamScape, investigates the role of multi-agent systems in enhancing narrative coherence, emotional richness, and multimodal creativity as part of AI-generated storytelling. Existing single-model story generators can lose logical coherence and cannot adapt in real-time to user interaction. To address the limitations of these existing systems, DreamScape introduces a cooperative multi-agent system that distributes the story generation process along an organized set of story dimensions (Plot, Character, Emotional, Visual, and Audio). The specialized agents communicate with one another through a team-based Coordinator Agent that can manage communication, memory, and conflict between agents using a structured set of negotiation protocols. The shared memory is maintained over the development of the story and is able to recall elements such as character, tone, and/a story event in order for the stories to maintain coherence across evolving story branches. It also integrates multimodal generation, providing both text, image, and audio generation in a synchronized manner to develop a richer sense of personal engagement in narrative experiences. This work develops connections between multi-agent coordination of LLMs, narrative intelligence, and human-AI co-creation as a growing area of research, to support applications in interactive entertainment (video games) education, and design of simulation systems. The system is constructed using fastAPI and is currently using ChromaDB and vector-based memory, with additional future extensions planned for narrative optimization based on reinforcement learning techniques.