MemFlow: Intent-Driven Memory Orchestration for Small Language Model Agents

Jiayi Chen, Yingcong Li, Guiling Wang · 2026

Read MemFlow: Intent-Driven Memory Orchestration for Small Language Model Agents by Jiayi Chen, Yingcong Li, Guiling Wang (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-2605-03312.txt.

Description

Jiayi Chen, Yingcong Li, and Guiling Wang present MemFlow, a memory system for small language model agents. Such agents must handle long, multi-turn histories. Full-context prompts overflow. Flat retrieval adds noise. Open-ended loops strain a small model's limited reasoning.

MemFlow plans memory outside the small model. A router sorts each query by intent. A memory agent then runs one of three tiers, from profile lookup to deep reasoning. Each tier has a matched token budget. A validator can retry with a heavier tier. On long-horizon benchmarks with a frozen small backbone, accuracy nearly doubled over full-context baselines.

Catalog record

CREATORJiayi Chen, Yingcong Li, Guiling Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03312

Browse the full library · About the Tower of Babel