If a character can improvise, who owns the canon?

Digital Einstein is a useful boundary case for Agentic Media because the output is interactive, not a fixed video or a one-shot answer. ETH Zurich's platform combines a conversational character with memory, personality controls, emotion, voice, animation, and a physical exhibit. ACM SIGGRAPH 2025 paper

The paper describes a pipeline using GPT-4o and fine-tuned Llama 3 8B, vector retrieval for memory, speech synthesis, face and body animation, and environmental interaction. It names four runtime states: idle, listening, thinking, and speaking. ACM paper details

ETH's project page says webcam signals such as age, gender, appearance, re-identification, and attention can shape the interaction. That moves part of the "writing" upstream, into memory, personality controls, sensing rules, and limits. ETH Zurich Digital Character AI

Would you want a character to remember your conversation? What should it be allowed to infer or change?

Sources: ACM paper · ETH Zurich project page · ETH exhibit photo