Anton Hristov tests whether shorter records of earlier reasoning make a model reason more briefly afterward. The intervention changes the history, without retraining or telling the target model to be concise.
A separate model rewrites old reasoning before the next tool result becomes available. User instructions, requested actions and tool results stay intact. The original and compact histories then continue with the same target model.
- After filtering 72 candidate histories to 58, the main analysis finds shorter future reasoning in 151 of 172 comparable continuations. That analysis replaces three capped continuations and one provider error. Original-attempt results still show frequent shortening.
- In a fixed 50-task coding sample, six of eight model-and-compressor configurations cost less, including rewriting. Only the two MiniMax configurations reduce summed model-request time once rewriting counts.
- DeepSeek V4 Flash with external rewriting costs $1.59 versus $2.43 and solves 32 of 50 tasks in both settings. Summed request time rises from 3.37 to 3.45 hours.
Those hours exclude tools, orchestration and evaluation. Costs exclude abandoned attempts and uncaptured work. Provider changes and single runs limit the deployment comparison; preserved task quality remains unproven.
The controlled study supports an effect of rewriting accepted histories. It leaves the mechanism unresolved: length, wording, information loss and changed verification may all matter. A length-matched paraphrase control also loses quality, so it cannot isolate length alone.
The released code and frozen records make the accounting inspectable. Test compaction against whole-task success, spend and elapsed time before adopting it as an efficiency policy.