context-compression

Context compression preserves critical state during long-running AI agent sessions by applying anchored summarization and artifact-trail tracking. It explicitly maintains structured sections for file paths, decisions, and continuation plans while trimming tokens to prevent forgotten changes.

Token budgeting for debugging continuity requires selecting a compression method like anchored, opaque, or regenerative summarization. This approach prioritizes total task cost over per-request savings, ensuring debugging state persists across truncated session histories without re-reading code.

Anchored summarization maintains a mergeable structured summary to reduce drift across repeated compression cycles, while regenerative context compression creates human-interpretable phase summaries. Opaque compression achieves maximum token reduction by sacrificing human readability for session needs.

Probe-based evaluation is needed to validate context compression quality when session history is truncated. It uses targeted recall, artifact, continuation, and decision probes to minimize hallucinated omissions and re-fetching, rather than relying on surface similarity metrics.

Context compression applies to coding, debugging, migration, and evaluation workflows where session history must be truncated. It preserves essential task state and artifacts, preventing the failure mode of losing track of key identifiers like function names and error messages during these processes.

Agents forget decisions after context truncation because standard summarization loses essential task state. Applying artifact-trail preservation with incremental merging explicitly tracks key identifiers and decisions, preventing failed continuation and the need to re-read modified files.