context-compression

To compress agent context and save tokens, apply anchored iterative summarization to reduce lengthy conversations while preserving essential state. This technique maintains persistent session context and file-level details to prevent information loss during compression.

Anchored iterative summarization is a state-management technique that compresses long-running agent conversations into structured summaries. It preserves critical files, decisions, and next steps across multiple compression cycles to maintain persistent session state without losing key artifacts.

To preserve decisions and files when summarizing code-debugging chats, generate structured summaries that explicitly retain artifacts, decisions, and next steps. This ensures critical debugging context survives the compression of lengthy conversation logs without losing key state.

To measure token savings versus information loss in context compression, use evaluation hooks to compare compressed structured summaries against original conversations. This validates whether critical decisions and files survive the summarization process effectively.

Use context compression for agent sessions when lengthy conversations approach context window limits and require efficient summarization. It is particularly applicable to long-running sessions and code-debugging chats where preserving state is critical for continued operation.

The best way to maintain persistent state across conversation compressions is using structured summaries with anchored iterative summarization. This approach ensures file-level context, decisions, and next steps survive each compression cycle without degradation.