Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation

Zhanghao Hu, Qinglin Zhu, Hanqi Yan, Yulan He, Lin Gui · 2026

Read Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation by Zhanghao Hu, Qinglin Zhu, Hanqi Yan, Yulan He, Lin Gui (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-2602-02007.txt.

Description

Zhanghao Hu and colleagues argue agent memory needs decoupling before aggregation. Agent memory forms a bounded tight stream with linked spans, so flat top k retrieval returns extra context. The team builds a revisable hierarchy of messages and groups. It segments history into local events, decoupling segments into components, aggregating related components into groups with a sparsity goal, maintaining it incrementally. At inference, xMemory retrieves top down, selecting a backbone of groups, then expanding when evidence helps. LoCoMo and PerLTQA tests show gains in quality and token efficiency.

Catalog record

CREATORZhanghao Hu, Qinglin Zhu, Hanqi Yan, Yulan He, Lin Gui
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63100
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.02007

Browse the full library · About the Tower of Babel