MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning (ICML 2026)

MemCast is a novel framework that reformulates Time Series Forecasting (TSF) as an experience-conditioned reasoning task. It explicitly accumulates forecasting experience into a hierarchical memory and leverages it to guide the inference process of Large Language Models (LLMs).

📝 “MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning”
ICML 2026 | 📄 Proceedings | PDF | arXiv | OpenReview

Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:119553–119569, 2026.

📰 News

• [2026-10-15] Publication update: MemCast is published in the ICML 2026 proceedings (PMLR 306, pp. 119553–119569). This README now includes the official proceedings entry, PDF, and conference BibTeX.
• [2026-10-15] Dataset update: Updated the dataset download link and preparation instructions, including data provenance and the chronological 8:2 memory/test split.
• [2026-02-03] Initial release: We introduce MemCast, a learning-to-memory framework for experience-conditioned time series forecasting. The first version of our paper is available on arXiv.

---

🙋 Please let us know if you find out a mistake or have any suggestions!

🌟 If you find this resource helpful, please consider starring this repository and citing our research.

Citation

Please cite the official ICML 2026 proceedings version:

@inproceedings{tao2026memcast,
  title={{M}em{C}ast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning},
  author={Tao, Xiaoyu and Cheng, Mingyue and Guo, Ze and Yu, Shuo and Liu, Yaguo and Liu, Qi and Wang, Shijin},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  pages={119553--119569},
  year={2026},
  volume={306},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  url={https://proceedings.mlr.press/v306/tao26a.html}
}

🔍 Overview

Existing LLM-based forecasting methods often treat instances as isolated reasoning tasks, lacking explicit experience accumulation and continual evolution. MemCast introduces a new learning-to-memory paradigm:

• Experience Accumulation: Organizes training data into a Hierarchical Memory consisting of historical patterns, reasoning wisdom, and general laws.
• Experience-Conditioned Reasoning: Retrieves relevant patterns to guide reasoning, utilizes wisdom for trajectory selection, and applies general laws for reflective iteration.
• Continual Evolution: Enables dynamic confidence adaptation to update memory entry confidence during inference without test data leakage.

---

✨ Key Features

• ✅ Hierarchical Memory: Structured storage of trend summaries, reasoning trajectories, and physical constraints.
• 🔗 Wisdom-Driven Exploration: Samples multiple reasoning paths and selects the optimal trajectory based on semantic consistency.
• 🔄 Rule-Based Reflection: Iterative refinement mechanism that enforces domain-specific constraints (e.g., non-negativity).
• 📊 Dynamic Adaptation: A confidence adaptation strategy that allows the model to evolve continuously in non-stationary environments.
• 🏆 SOTA Performance: Consistently outperforms deep learning and LLM-based baselines on Energy, Electricity, and Weather benchmarks.

---

🚀 Getting Started

1. Clone the repo

git clone https://github.com/Xiaoyu-Tao/MemCast-TS
cd MemCast

2. Environment Setup

conda create -n memcast python=3.10
conda activate memcast
pip install -r requirements.txt

3. Prepare Data

MemCast is evaluated on diverse real-world datasets with rich contextual features: Energy (NP, PJM, BE, FR, DE) Electricity (ETTh1, ETTm1) Renewable Power (Windy Power, Sunny Power) Hydrology (MOPEX) Download the datasets used in our experiments from Google Drive (memcast_dataset.zip) and extract the 10 CSV files into ./datasets/.

The datasets follow those used in AlphaCast. ETTh1 and ETTm1 are manually cropped subsets of the official ETT datasets; no additional preprocessing script is provided. MemCast uses a chronological 8:2 split: the first 80% of each dataset is used for memory accumulation, and the final 20% is used for testing. No separate validation split is used.

mkdir -p datasets
unzip /path/to/memcast_dataset.zip -d datasets/

4. Experience Accumulation (Memory Construction)

sh scripts/accumulation/build_memory.sh

5. Short-term Forecasting (e.g., NP)

sh scripts/short_term/NP.sh

6. Long-term Forecasting (e.g., ETTh)

sh scripts/long_term/ETTh.sh

7. Evaluation

After running forecasting, evaluate prediction quality using the uqmem scoring pipeline (LLM + heuristic trajectory selection):

Short-term (EPF / NP, look_back=168, pred_window=24):

bash scripts/evaluate/EPF.sh

Long-term (ETTh1, look_back=96, pred_window=96):

bash scripts/evaluate/ETTh.sh

Or run the evaluation scripts directly:

# EPF
python evaluate/EPF/evaluate_EPF_uqmem.py

# ETTh
python evaluate/ETTh/evaluate_ETT_uqmem.py

The evaluation reports MSE and MAE metrics for each attribute and prints a per-sample summary.

---

📊 Benchmark Results

Full Results: Overall Performance: MemCast consistently achieves the best or second-best results across datasets, surpassing Time-LLM and TimeReasoner.

Ablation Results:

 

Ablation Study: Validation of memory components and dynamic adaptation strategy.