context-compression

Use anchored iterative summarization: define explicit summary sections for session intent, files modified, and decisions, then summarize only newly truncated content on each compression trigger and merge it into the existing summary. This prevents the drift that occurs when summaries are regenerated from scratch.

Anchored iterative summarization scores highest at 3.70 overall with 98.6% compression, compared to 3.44 for regenerative summaries and 3.35 for opaque compression. It performs best on accuracy and context awareness, which matter most for coding tasks.

Artifact trail is the weakest dimension across all compression methods, scoring only 2.2-2.5 out of 5. General summarization cannot reliably track file operations, so implement a separate artifact index or explicit file-state tracking in the agent scaffolding.

Default to a sliding window approach that keeps the last N turns plus a structured summary, or trigger at 70-80% context utilization. Use task-boundary triggers when sessions have clear phase transitions like research, implementation, and testing.

Use probe-based evaluation instead of metrics like ROUGE. After compression, ask recall, artifact, continuation, and decision questions, then score responses with an LLM judge across six dimensions including accuracy and artifact trail.

Never compress tool definitions, function call schemas, or API specs, since summarizing parameter names or types destroys the agent's ability to invoke tools. Also protect early session turns containing task constraints and preserve code blocks and identifiers verbatim.