An alternative engineering view treats the boundary as negotiable. External memory, retrieval systems, adapters, activation patches, or periodic fine-tuning can turn temporary information into a durable state without rewriting the base model. Under this definition, the relevant system is the model-plus-memory stack, not the frozen neural network alone. The advantage is practical persistence; the cost is added machinery, provenance risk, retrieval failure, and a less clean claim that the model itself has learned.