The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents

Jiayuan Liu 1,2 , Tianqin Li ∗1 , Shiyi Du ∗1 , Xin Luo ∗3 , Haoxuan Zeng 3 , Emanuel Tewolde 1,2 , Tai Sing Lee 1 , Tonghan Wang 4 , Carl Kingsford 1 , Vincent Conitzer 1,2 1 Carnegie Mellon University 2 Foundations of Cooperative AI Lab (FOCAL) 3 University of Michigan 4 Harvard University {jiayuan4, tianqinl, shiyid}@cs.cmu.edu , {luosanj, zchx}@umich.edu , [email protected] , [email protected] , [email protected] , [email protected] , [email protected] · 2026

Read The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents by Jiayuan Liu 1,2 , Tianqin Li ∗1 , Shiyi Du ∗1 , Xin Luo ∗3 , Haoxuan Zeng 3 , Emanuel Tewolde 1,2 , Tai Sing Lee 1 , Tonghan Wang 4 , Carl Kingsford 1 , Vincent Conitzer 1,2 1 Carnegie Mellon University 2 Foundations of Cooperative AI Lab (FOCAL) 3 University of Michigan 4 Harvard University {jiayuan4, tianqinl, shiyid}@cs.cmu.edu , {luosanj, zchx}@umich.edu , [email protected] , [email protected] , [email protected] , [email protected] , [email protected] (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-2605-08060.txt.

Description

Jiayuan Liu, Tianqin Li, Shiyi Du, Xin Luo, and coauthors report a counterintuitive failure of language model agents. Giving an agent a longer memory of past rounds is usually treated as a pure upgrade. Across seven models and four social dilemma games played over five hundred rounds, longer recall instead reduced cooperation in most settings, a pattern the authors call the memory curse.

Analysis of hundreds of thousands of reasoning traces links the decay to fading forward-looking intent, not rising suspicion. Replacing visible history with synthetic cooperative records restored cooperation at the same prompt length, showing that memory content, not length, drives the effect. Targeted fine-tuning on forward-looking traces also reduced the decay.

Catalog record

CREATORJiayuan Liu 1,2 , Tianqin Li ∗1 , Shiyi Du ∗1 , Xin Luo ∗3 , Haoxuan Zeng 3 , Emanuel Tewolde 1,2 , Tai Sing Lee 1 , Tonghan Wang 4 , Carl Kingsford 1 , Vincent Conitzer 1,2 1 Carnegie Mellon University 2 Foundations of Cooperative AI Lab (FOCAL) 3 University of Michigan 4 Harvard University {jiayuan4, tianqinl, shiyid}@cs.cmu.edu , {luosanj, zchx}@umich.edu , [email protected] , [email protected] , [email protected] , [email protected] , [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE106 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.08060

Browse the full library · About the Tower of Babel