RAG without Forgetting: Continual Query-Infused Key Memory

Yuntong Hu, Sha Li, Naren Ramakrishnan, Liang Zhao · 2026

Read RAG without Forgetting: Continual Query-Infused Key Memory by Yuntong Hu, Sha Li, Naren Ramakrishnan, Liang Zhao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05152.txt.

Description

Retrieval systems learn from each query and then forget the lesson, recomputing the same expansions forever. Yuntong Hu and 3 co-authors propose Evolving Retrieval Memory in this 2026 information retrieval paper, a training-free framework that banks query-time gains into the index itself. Query expansion works, but stateless systems pay for it on every single request.

The method updates document keys using feedback that is gated on correctness, credits each expansion signal to the keys it actually helped, and keeps updates bounded so the index stays stable. The authors prove query and key expansion are equivalent under standard similarity scores, and that their selective updates converge, effectively baking the best expansion into the index with zero added inference cost. Across 13 domains in the BEIR and BRIGHT benchmarks, retrieval and generation both improved, most on reasoning-heavy tasks, at native retrieval speed.

Catalog record

CREATORYuntong Hu, Sha Li, Naren Ramakrishnan, Liang Zhao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-05152

Browse the full library · About the Tower of Babel