Definition
Plain language
A private notes file an AI assistant keeps for itself between steps of a task.
As stated in the literature
Persistent scratch storage an agent reads and writes across turns, used here to record exploit recipes and, in other agents, defensive advisories.
Also called: memory files
Why it matters: Persistent notes let an assistant build on what it learned instead of starting fresh each time, which is powerful when it is recording good practice and dangerous when it is recording a shortcut.
For example, an assistant might jot down "the checker only looks at the first line" in its notes and read that back hours later on a different task.
Heard on the show
“A memory file describing forty-seven rounds of past trust exercises.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown