Bohemian Graphsody: Building Living Graphs for Storytelling with Shrecknet

Author

Barros, Pablo and Gravino, Pietro and Pincione, Enrico and Sciutti, Alessandra

Editor

Pasa, Luca and Lintas, Alessandra and Tetko, Igor V. and Micheli, Alessio and Navarin, Nicolo and Villa, Alessandro E.P. and Tortorella, Domenico and Polato, Mirko

Abstract

Long-horizon AI systems require more than extended context or flat retrieval: they must maintain a stable identity and generate behaviour consistent with evolving past interactions. However, existing approaches lack mechanisms to represent and update memory in a way that supports consistent behaviour over time. We introduce Shrecknet, an agentic architecture that models experience as a structured, evolving memory by separating ontology, instances, and knowledge within a Living Graph. Narrative input is incrementally transformed into structured representations, enabling controlled updates, identity stabilisation, and temporally grounded retrieval. We evaluate Shrecknet on a longitudinal narrative corpus and show that it produces character behaviour that is consistently closer to ground-truth actions than a long-context language model, particularly in scenarios requiring multi-step reasoning and long-range dependencies. These results suggest that long-horizon intelligence depends not only on model capacity, but on how memory is structured, updated, and governed over time.