Definition
Plain language
A separate model that watches an AI agent's running notebook and rewrites it into a clean summary when it gets too full.
As stated in the literature
A dedicated module in long-horizon agent systems that compresses accumulated rollout context into structured state to keep working memory within budget without losing epistemic state.
Why it matters: It lets long-horizon agents keep working coherently past the point where their raw history would overflow the context window.
For example, a side model watches a research agent's scratchpad and, when it nears its limit, rewrites it as 'tried approaches A and B (failed), currently exploring C with hypothesis X.'
Heard on the show
“Yeah, the context condenser is my favorite engineering move in the paper.”Episode 082 — Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick