Chandra Vamsi Krishna Alla

BudgetMem: Training-Free Selective Memory for Cost-Efficient Long-Context Processing in Language Models

BudgetMem is introduced, a training-free architecture that keeps only high-salience content under an explicit memory budget that makes chunk-level keep-or-discard decisions from interpretable features: entity density, TF-IDF importance, position, numerical density, discourse markers, and question presence.