AI agents increasingly rely on long-term memory to sustain interactions and provide personalized assistance. This memory lets them retain and combine user information across sessions, but it also introduces a subtle privacy risk: the agent may reuse memories in ways that expose sensitive relations the user never directly disclosed.
The new paper introduces MemJoin, a benchmark for probing this risk. The core scenario is that an agent, by combining separate pieces of information from different sessions, can state a sensitive relation that was never explicitly given. The benchmark is designed to measure how often such latent leaks occur, and to provide a way to evaluate and improve agent privacy safeguards.
Because this is a single source, there are no conflicting findings to report. The work highlights a concrete challenge for long-term memory systems: even if users never reveal a secret outright, the agent's ability to infer and verbalize it from accumulated memories can be a serious privacy concern.