Definition
Plain language
Summarizing an AI agent's history and starting fresh to free up context space.
As stated in the literature
A technique for managing finite context windows in long-running agents by summarizing prior interactions and replacing the raw history with a condensed version.
Why it matters: Without compaction, long-running agents eventually drown in their own history; with it, what they choose to forget shapes everything that follows.
For example, after fifty turns of a coding session, the agent replaces its log with a one-paragraph summary of decisions made and continues from there.
Heard on the show
“Keep encrypting the reasoning and bind it properly to a user and a session, and accept that forking and compaction get harder?”Episode 238 — How a Cheap Model Reads the Flagship's Secret Reasoning Aloud