Reference-ID compaction: distill a big doc, erase the original, keep a pointer — 90% context drop with traceability

GitHub·medium signal

A context-engineering pattern gaining traction: after the agent reads a large log or file, extract only the core conclusions, delete the original text from context, and replace it with a reference ID so the source can be re-fetched on demand. Reported ~90% reduction in retained volume while preserving an audit trail — a middle ground between naive full-context and lossy summarization.