A Temporal-Error State Architecture for Long-Horizon AI Agents

Abstract

This record contains a research architecture proposal and evaluation protocol for the Temporal-Episodic Agent Architecture (TEAA), focused on long-horizon AI agents that must reason under deadlines, estimate task completion time, and adapt after execution failures. The proposed architecture introduces a candidate Continuous Temporal-Error State (CTES) module intended to represent elapsed time, remaining deadline, estimated task progress, predicted completion time, and prediction error. A separate episodic memory store preserves timestamped events and outcomes, while a retrieval controller selects relevant past experiences for use by a large language model (LLM)-based agent. The manuscript defines research questions, candidate mathematical formulations, baseline systems, ablation studies, evaluation metrics, and safety considerations. Its central hypothesis is that a trainable temporal-error state may improve deadline-sensitive task performance compared with prompt-based time injection and simpler recurrent-state baselines under matched computational budgets. Research status: This is a research proposal and evaluation protocol. It does not report completed experiments or validated performance improvements. The novelty and practical benefits of the proposed mechanism remain to be established through implementation, systematic comparison with prior work, and reproducible experiments.