Fine-tuning Image Transformers using Learnable Memory