Preserve the goal, decisions taken and why, unresolved threads, and identifiers such as file paths and record ids. Discard the intermediate reasoning that led to a settled decision, which is the bulk of the tokens and almost never load-bearing again.

University of Michigan - Ann Arbor

A topic of the Building AI Agents course. Planned for 1.5 hours at Hard difficulty.

Context Compaction and Summarisation

Compaction replaces a long stretch of transcript with a shorter summary that preserves what later steps will need. It is what lets an agent run for hours in a context window that would otherwise have overflowed within the first twenty minutes.

Learn Before

Related