noun · AI technique
Compaction
also called context summarization
Squashing the old pages into a short note when the backpack gets full, so there is room to keep working.
Replacing earlier conversation content with a shorter summary to free context space in a long session.
When a conversation nears the context window limit, the harness asks a model to summarize the older part and replaces it with that summary. The details in the original are gone from then on, and only the summary is resent.
Real-life example
You work with a coding assistant for two hours. Near the limit, the app quietly summarizes the first hour into a paragraph. Later you notice it has gone back to the old naming style you told it to avoid. Your instruction was in the part that got squashed, and the summary did not keep it.
Conversation compacted: 142 messages summarizedWhere the analogy breaks
-
Compaction loses detail. A summary keeps the gist, and a small but important instruction can fall out of it.
-
You may not control when it happens. Many tools do it automatically near the limit, so important rules are safer written in a file that is reloaded each time.
-
It is not the same as dropping old messages. Dropping removes them; compaction keeps a short version, which can also be wrong.