context-compression

Context compression minimizes tokens in long-running agent sessions by preserving critical signals while trimming noise. Use anchored iterative summarization to merge new summaries with a persistent anchor, producing compact per-task summaries that balance fidelity against reconstruction cost.

Artifact tracking explicitly logs created, modified, and read files within compressed contexts. This prevents debugging cases where file changes are forgotten after summarization, maintaining visibility into agent state across session boundaries.

Yes. Structured, persistent summaries with explicit sections enable context compression while retaining essential conversation signals. Three production-ready variants—anchored iterative summarization, opaque compression, and regenerative full summarization—offer different trade-offs between compression ratio and reconstruction fidelity.

Apply compression when agents face context window limits, designing conversation summarization strategies or evaluating compression approaches. It's essential for debugging forgotten modifications and building evaluation frameworks for compression quality in coding or debugging sessions.

Opaque compression encodes context densely without explicit structure, minimizing size at the cost of reconstruction complexity. Regenerative full summarization rebuilds complete context on demand, prioritizing fidelity and debuggability over compression ratio in trade-off scenarios.