This presentation was delivered as an invited talk at Sogang University in August 2026.
What changes when lifelong learning is viewed as a problem of skill memory?
An embodied agent that learns after deployment must do more than acquire new behaviors. It must decide what knowledge to update, when an existing skill can be reused or combined with another, and what information should be retained for future adaptation. A skill-memory perspective therefore shifts the focus from repeatedly retraining a complete policy to maintaining a structured collection of reusable skills.
This presentation develops that perspective through three questions:
- Locality — How can the agent update only the knowledge affected by new experience while preserving unrelated capabilities?
- Compatibility — How can it determine whether existing skills can be reused or combined under the current initial and target conditions?
- Efficiency — What is the minimum information the agent must retain so that future adaptation does not require relearning or storing an entire policy?
Together, these distinctions change how lifelong learning is designed: learning becomes selective updating, action becomes compatibility-aware reuse and composition, and memory becomes compact support for future adaptation.
The presentation is interactive. Use the arrow keys to navigate, and press Space on demonstration slides to play their animations. On a phone, rotate to landscape, swipe between slides, and tap a demonstration slide to play it.