context-compression

Compress long agent conversations using anchored iterative summarization to condense session histories into structured, compact representations while preserving essential signals like facts, decisions, and file changes to minimize token usage.

Anchored iterative summarization for agent memory maintains persistent summaries with explicit sections for intent, files modified, decisions, and next steps to preserve critical signals in memory for continuity across compression cycles.

Summarize a multi-turn debugging workflow without losing context by generating structured, sectioned reports that capture intent, modified files, decisions, and next steps while minimizing token usage.

Evaluate context compression quality for agent sessions using a lightweight evaluation framework with probe-based scoring and artifact-trail awareness to ensure critical signals are preserved during compression.

You need to compress context for large code exploration histories when sessions exceed context windows, requiring tokens-per-task optimized representations to maintain continuity without losing critical signals.