Definition
Plain language
A trick that lets a forgetful AI agent 'remember' its past by re-feeding it its entire history of steps on every new turn.
As stated in the literature
In agent scaffolds, the practice of re-sending the full accumulated trajectory (commands, outputs, edits) into a stateless model on every step; replay cost grows with history length and can dominate token usage on long-horizon tasks.
Why it matters: It is the simple workaround that gives a memoryless model a sense of history, but its cost balloons on long tasks and can swallow most of the budget.
For example, an agent fixing a bug across fifty steps re-reads every command and output from steps one through forty-nine before taking step fifty, so its reading load keeps growing.
Heard on the show
“That's called context replay.”Episode 125 — AI Coding Agents Run a Marathon, and Fewer Than One in Three Finish