memleak-bench/

memleak-benchmark

MemLeak is a benchmark for diagnosing information leaks in multimodal agent memory systems. It measures whether AI agents truly forget user information after deletion requests, or whether retained images allow VLMs to reconstruct deleted facts through implicit visual features.

	Dataset Description




	
	Overview

113 synthetic entity profiles (20 facts each, 5 categories: location, profession, preference, relationship, possession) 536 synthetic… See the full description on the dataset page: https://huggingface.co/datasets/memleak-bench/memleak-benchmark.