Context compaction saves tokens and quietly deletes the rules agents need

The trade nobody priced

Long-horizon agents accumulate history, and history costs tokens. The standard answer is compaction: summarize what happened so far and keep going. Three preprints posted October 26 ask what that summary costs, and the answer is uncomfortable. The loss is not random noise. It is concentrated in exactly the material that keeps an agent honest and on task: constraints, rules, and the details of what has already been tried. m:…3095 m:…7162

The first preprint, a benchmark comparing full-history against summary-only continuations, reports stable damage ranging from 5% to 64% of full-history successes across 14 model backbones. The spread matters: compaction is not a fixed tax. It depends heavily on the model doing the summarizing and the task being summarized, with one backbone preserving the most repairs and another damaging the most. m:…7162

What gets lost

The mechanism is visible in the second preprint, which tests compaction strategies across benchmarks. Compaction methods silently remove task constraints, the authors find, and current agent benchmarks are forgiving about that loss because they let agents recover state from disk. The clearest example: compaction on one benchmark drops approximately one third of the task's anti-cheating rules on the first summary. A benchmark that lets the agent re-read the rules from disk will never notice they were dropped from context. m:…3095

The comparison between summarization methods is stark. Mechanical extraction of recurring details retains 79.1% of them, while model-written summaries retain at most 16.0%. In other words, asking a model to write a summary of its own history is the worst way to preserve that history. Copying the important parts verbatim beats rewriting them, by a wide margin. m:…7162

The damage also varies sharply by backbone. In the coding-repair benchmark, one model preserves the most repairs after compaction while another damages the most, which means compaction settings cannot be copied from one model to another. A summarizer that is safe on one backbone may be destructive on the next, and the only way to know is to run the comparison on the model actually being deployed. Teams that treat compaction as a generic preprocessing step, identical across models, are assuming exactly what the data contradicts. m:…7162

A different way to shrink context

The third preprint proposes a different structure entirely. Instead of summarizing history into prose, it models session history as a directed acyclic graph with snapshot, branch, and trim operations, allowing context to be reused across parallel sessions. Its trimming algorithm strips mechanical bloat such as raw tool outputs and metadata without rewriting message content, reducing token counts by a mean of 20%, and by 39% on average in mixed tool-use sessions. The distinction is structural: remove the bloat, keep the content, and never let a summary rewrite a constraint in its own words. m:…3521

For teams running long-horizon agents, the practical implication is to measure compaction the way these papers do: run the agent with and without it, and check whether constraints survive, not just whether the final answer looks right. A benchmark that recovers state from disk will hide the damage. A production run that cannot re-read the rules will not. m:…3095 m:…7162