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

TS-Memory addresses a critical bottleneck in the deployment of Time Series Foundation Models (TSFMs) like Google's TimesFM or Amazon's Chronos: the trade-off between expensive RAG-style retrieval at inference time and the risks of catastrophic forgetting during fine-tuning. While the project presents a clever distillation-based approach to bake non-parametric knowledge into a parametric memory module, its current defensibility is low (0 stars, though 9 forks suggest active research interest). The moat is purely academic/algorithmic; there is no network effect or data gravity. The high platform domination risk stems from the fact that providers of the underlying TSFMs (Google, Amazon, Salesforce) are likely to implement their own native 'memory' or 'context window' optimization techniques directly into their managed APIs. This project is currently a reference implementation of a paper rather than a production-ready tool, making it easily reproducible by any ML engineering team working on forecasting pipelines.