Definition
Plain language
Deleting a particular thing the AI had written down, then seeing how its behavior changes.
As stated in the literature
Targeted editing or removal of specific stored claims from an agent's persistent memory files, followed by repeated re-evaluation, to establish a causal link between a remembered technique or belief and downstream task performance.
Also called: memory interventions, memory edit, memory edits
Why it matters: It turns a correlation — the agent wrote something down and later got better — into evidence that the written note was actually doing the work.
For example, researchers delete the single line where the agent wrote down that jumping while throwing sends objects higher, then re-run the throwing task to see if its success rate collapses.
Heard on the show
“In the third memory intervention, deleting it improves the agent's tallest-tower score.”Episode 288 — An AI Agent Given Thirty Hours and No Goal, Then Tested on What It Learned