A number of transformations in high-level MLIR dialects like Linalg, SCF and Affine focus on building transformations for optimizing the cache behavior of programs. While building models for the cost of computation for these transformations are easy as it is a local property, data movement on cached architectures depends on the global state and is very hard to predict. Current approaches for memory models for these transformations use transformation-specific cost models, which do not compose from one transformation to another due to the global properties of data movement. We introduce mlir-meminfo, a lightweight, analytical memory model for MLIR. mlir-meminfo accurately predicts the cache behavior of a program, with a particular focus on transformations, updating the cache information almost instantly for modern neural networks like BERT, containing hundreds of memory accesses. With the introduction of mlir-meminfo, we aim to revolutionize transformation memory models for MLIR and improve developer productivity for performance programming.