CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory

Official implementation of CREAM, a self-supervised framework for memory-based continual retrieval on dynamic streaming corpora.
CREAM learns continual retrieval in a fully unsupervised setting with adaptive soft memory.

TL;DR

CREAM proposes soft memory for practical continual IR under unbounded / unlabeled / topic-shifting streaming corpora, and introduces three key techniques:

1. Fine-grained similarity (token-level semantics)
2. Regularized cluster prototypes (fixed token length prototypes via LSH-style regularization)
3. Stratified coreset sampling (diverse training samples from memory)

Method Overview

CREAM repeats three stages per streaming session:

• Retrieval: return relevant documents with the up-to-date encoder
• Memory Update: streaming clustering with regularized prototypes (soft memory maintenance)
• Encoder Update: self-supervised training via contrastive objective using pseudo pos/neg sampled from memory
  

Datasets

• LoTTE
  • https://huggingface.co/datasets/colbertv2/lotte
  • https://huggingface.co/datasets/colbertv2/lotte_passages
• MSMARCO
  • https://github.com/naver/ms-marco-shift

Quick Start

1. Generate sessions(cream/src/data)

# Raw sessions
python generate_multi_test.py

# Filtering with BM25
python proposal_input_helper.py

2. Train and evaluate(cream/src)

• You can evaluate with rolling evaluation optionally.

python main.py \
 --exp=proposal_qq_low \
 --use_tensor_key \
 --warming_up_method=stream_seed \
 --sspq=50 \
 --start=0 \
 --end=10 \
 --rdsz=50 \
 --cmnsz=50 \
 --mi=3 \
 --init_k=5 \
 --light_weight \
 --light_weight_rate=0.25

Citation

@inproceedings{son2026cream,
  title     = {CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory},
  author    = {Son, HuiJeong and Kang, Hyeongu and Kim, Sunho and Ho, Subeen and Kang, SeongKu and Lee, Dongha and Yoon, Susik},
  booktitle = {Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '26)},
  year      = {2026}
}