A startup memory loader is the part of a system that reads saved instructions, preferences, or session memory when a tool or agent begins running. In AI assistants and other automation software, this code often runs before the user has reviewed what will be loaded, so it becomes a security-sensitive trust boundary.
It matters because a loader can accidentally import data from outside the intended project or workspace if it follows paths without validating their real destination. A common risk is symlink handling: the path may look harmless, but the resolved target can point somewhere else. Attackers may use this to expose private files to an assistant’s context, while defenders harden loaders by canonicalizing paths, checking the resolved location against an allowed root, and limiting automatic loading to least-privilege sources.