context-compression

Compress conversation history by applying anchored iterative summarization to retain goals, key decisions, and modified artifacts while trimming verbose context. This reduces token usage while preserving essential information for long-running agent sessions.

Preserve modified artifacts during context compression by using structured artifact tracking. The session state is sectioned into intent, files, decisions, and next steps to maintain artifact integrity across compressions.

Validate information integrity after summarizing by prompting the agent with recall, artifact, continuation, and decision probes. These evaluation probes check the compressed context against the original history to ensure critical details remain intact.

Yes, context compression works for multi-file project discussions. It applies structured session state management to track modified files and decisions, preventing performance degradation when history growth threatens the session.

Apply context compression in LLM evaluation workflows when a session accumulates thousands of tokens of history. It trims verbose context while retaining structured state for reproducibility and continued evaluation.