Instruction-Following LLM Agent for Simulating Human Memory Retrieval Process in Household Tasks

Abstract

Human decision-making in dynamic household environments depends on memory retrieval processes that shape action selection and timing. This paper presents an instruction-following LLM agent as an executable human performance model for simulating memory retrieval behavior in household tasks. The proposed framework integrates memory encoding and decision-making modules with a modified ACT-R base-level activation equation that incorporates environmental entropy and observation duration. The framework was evaluated using human behavioral data from an AI2-THOR-based household task study. When predicting human search times, the modified equation produced lower prediction errors than the conventional ACT-R baseline, reducing MAE, RMSE, and MAPE by 11.08%, 5.37%, and 16.00%, respectively. These findings provide preliminary evidence that incorporating context-sensitive retrieval factors may support closer alignment with human search-time behavior in executable LLM-based human performance models.