Philosophically, the project complicates the usual distinction between memory and learning. Retrieval-based memory preserves an external record, while JIT-LoRA attempts to sediment interaction into parameters. That makes personalisation more immediate, but also less transparent: a model may behave differently without exposing a clean, inspectable memory trace. The central issue is not whether adaptation is possible. It is whether an agent should be permitted to change its dispositions during dialogue, and how those changes can be audited, reversed, or attributed.