Definition
Plain language
A running list of bad edits an AI was about to make so it doesn't try the same mistake again.
As stated in the literature
In SkillOpt, a persistent training-time memory of edits rejected by the validation gate (with associated score drops), shown in-context to the optimizer on subsequent rounds to prevent re-proposing failed changes.
Why it matters: Without a memory of past failures, the optimizer keeps re-suggesting the same bad ideas; the buffer makes self-improvement actually move forward rather than oscillate.
For example, if a proposed change to the skill document hurt validation score last round, that exact edit is shown to the optimizer next round labeled "do not propose again."
Heard on the show
“Which is — Juniper, this is your favorite part, I think — the rejected-edit buffer.”Episode 078 — Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training