Long-horizon LLM agents often exceed their context windows during extended interactions. The standard fix is context compaction, where the agent's interaction history is replaced with a natural-language summary. But a new arXiv preprint argues that these summaries contain more than a record of what happened, making the compaction step itself a bottleneck.
The paper, Prune the Summary, Not the History, proposes source-judged compaction. Instead of pruning the raw history, the method prunes the summary, using the original source to judge which summary content is worth keeping. This inverts the usual pipeline: the summary is treated as the object to be trimmed, not the history.
If the approach holds up, it could make long-horizon agents more reliable by reducing the information loss that comes from summarization. The authors identify summary-driven compaction as a key limiting factor for agents that need to remember early steps in a long task.