Adaptive Memory Admission Control for LLM Agents

Guilin ZhangWei JiangXiejiashan WangAisha Behr, Kai ZhaoJeffrey FriedmanXu ChuAmine Anoun [2pt] Workday AI · 2026

Read Adaptive Memory Admission Control for LLM Agents by Guilin ZhangWei JiangXiejiashan WangAisha Behr, Kai ZhaoJeffrey FriedmanXu ChuAmine Anoun [2pt] Workday AI (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-2603-04549.txt.

Description

Zhang and colleagues control what language model agents remember with Adaptive Memory Admission Control. Agents lean on long term memory for multi-session reasoning, yet current systems barely govern retention, hoarding stale or hallucinated facts or relying on opaque, costly policies. A-MAC treats admission as a structured choice, scoring content type prior, temporal recency, semantic novelty, factual confidence, and future utility. Lightweight rules extract features, one model assisted assessment judges utility, and cross-validated optimization learns adaptive policies, reaching F1 of 0.583 on LoCoMo with latency down 31 percent.

Catalog record

CREATORGuilin ZhangWei JiangXiejiashan WangAisha Behr, Kai ZhaoJeffrey FriedmanXu ChuAmine Anoun [2pt] Workday AI
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28416
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.04549

Browse the full library · About the Tower of Babel