TS-Memory: Plug-and-Play Memory for Time Series Foundation Models

The article introduces TS-Memory, a lightweight memory adapter designed to enhance Time Series Foundation Models (TSFMs) without requiring significant changes to the model architecture. TS-Memory employs a two-stage training process that involves creating a kNN teacher for generating quantile targets and distilling this information into a memory adapter, allowing for efficient, retrieval-free inference. This approach addresses the challenges of catastrophic forgetting and high inference latency, providing significant improvements in forecasting accuracy while maintaining efficiency comparable to the original frozen models, which is crucial for practitioners dealing with distribution shifts in time series data.