Peter Belcak and colleagues at NVIDIA test whether AI assistants can work from short handoff notes instead of retaining every earlier message. They compare full histories with periodic summaries and with a fresh conversation after every action. Large source documents remain searchable files.
The strongest results concern two proprietary models, Claude Opus 4.5 and GPT 5.6 Sol. The study covers six multi-step task suites and seven long-document suites. These are the authors’ reported experiments, not independent replications.
- On 500 software-repair tasks, summarizing after every action gives Opus 70.2% success versus 70.0% with full history. GPT falls to 82.4% from 84.4%. Median peak conversations shrink to roughly 7,000–9,000 tokens, the units models process and generate.
- Cost reductions of 2.2–8.8 times for Opus and 2.4–5.0 times for GPT assume hypothetical context-sensitive prices. The authors derive them from serving measurements on different, downloadable models.
- At standard published prices, GPT’s software-repair run costs $203.12 versus $167.68 with full history when summaries start at 12,000 tokens within a 16,000-token limit. More generated text can outweigh shorter history.
- Tight summaries transfer unevenly to local models. Qwen 3 Coder Next drops from 53.0% to 3.6% on software repairs at that same limit.
Document-search gains also need a separate explanation. In the hidden-fact retrieval and scattered-number tests, the models improve without performing any summaries. Searching the files, rather than repeated summarization, explains those reductions in working history.
The practical design is a compact execution brief: preserve the task, verified findings, changed files, remaining steps and completion checks. Keep original evidence retrievable. The paper proposes training models specifically for this loop, but does not test such training.
Caveat: These fixed-task experiments do not establish reliable memory under changing human instructions, universal success across local models, or the advertised savings on current API bills.