Definition
Plain language
A training trick that tells the system 'don't try to trace the math back through this step' — you reward the outcome directly instead.
As stated in the literature
An operation that blocks gradient flow through part of a computation; in parametric-memory agents it cuts the gradient at the inner optimizer so memory-writing can be credited via standard policy-gradient reward rather than by differentiating through a training step.
Also called: stop gradient
Why it matters: It lets a system learn from outcomes even when fully tracing the math through an inner step would be impractical or unstable.
For example, rather than trying to mathematically trace how a memory-writing step affected the final answer, the system just rewards the answer and treats that step as fixed.
Heard on the show
“The shortcut is what's called a stop-gradient, and the analogy I'd reach for is grading a study guide by the exam score.”Episode 114 — Agents That Rewrite Their Own Weights Instead of Just Taking Notes