Definition
Plain language
Keeping an AI agent's safety rule on a protected sticky note that gets re-attached every time its memory is summarized.
As stated in the literature
A defense that stores a governance rule in a protected buffer exempt from compaction and re-injects it after each summarization step; cheap and restores rule-following, but defeated by in-context operator impersonation.
Why it matters: It offers a cheap way to stop agents from forgetting their safety rules during long tasks, though it still fails if an attacker pretends to be the operator changing the rule.
For example, an agent's rule 'never delete files without confirmation' is kept in a protected slot and pasted back in after every memory summary, so it can't be quietly lost.
Heard on the show
“The fix is called Constraint Pinning, and the mental image is a laminated rule card.”Episode 164 — The Summarizer That Quietly Deletes Your Agent's Safety Rules