EMemBench: Interactive Benchmarking of Episodic Memory for VLM Agents

TL;DR

EMemBench is a new interactive benchmarking framework for evaluating episodic memory in vision-language agents through game-based questions, revealing persistent challenges in visual memory tasks and the need for further research.

Contribution

Introduces EMemBench, a novel interactive benchmark for episodic memory in agents, with question generation from agent trajectories and comprehensive evaluation across text and visual environments.

Findings

Memory performance varies significantly across tasks.

Visual memory remains a challenging area for current models.

Human study confirms the benchmark's difficulty.

Abstract

We introduce EMemBench, a programmatic benchmark for evaluating long-term memory of agents through interactive games. Rather than using a fixed set of questions, EMemBench generates questions from each agent's own trajectory, covering both text and visual game environments. Each template computes verifiable ground truth from underlying game signals, with controlled answerability and balanced coverage over memory skills: single/multi-hop recall, induction, temporal, spatial, logical, and adversarial. We evaluate memory agents with strong LMs/VLMs as backbones, using in-context prompting as baselines. Across 15 text games and multiple visual seeds, results are far from saturated: induction and spatial reasoning are persistent bottlenecks, especially in visual setting. Persistent memory yields clear gains for open backbones on text games, but improvements are less consistent for VLM agents, suggesting that visually grounded episodic memory remains an open challenge. A human study further confirms the difficulty of EMemBench.