context-compression

To compress long agent conversations, apply anchored iterative summarization to merge new content into structured sections and preserve critical context while reducing tokens.

Anchored iterative summarization is a context compression method that maintains a persistent, sectioned summary, merging new content into existing structures to prevent information drift.

Prevent silent information loss during large codebase summarization by using explicit artifact tracking to maintain an index of files read, modified, and created throughout the session.

Yes, you can use opaque compression to achieve very high token savings for low fidelity requirements, trading interpretability to reduce context size without losing critical meaning.

The best way to summarize debugging tasks within token limits is a three-phase approach: capture session intent, compress by merging content into structured sections, then integrate only new spans.

Avoid regenerative full summary when low fidelity is acceptable or token limits are strict, because recomputing detailed summaries on each cycle is resource-intensive and reduces token savings.