Octopus preprint reports gains from sharing memory across AI agents

Researchers report in a preprint that Octopus, a framework for coordinating AI agents within one model, reached 49.6% accuracy on BrowseComp-Plus. That was 17.5 percentage points above single-agent reinforcement learning at 143K context and 12.5 points above multi-agent reinforcement learning using text summaries. Octopus lets branches share stored attention states on a GPU and read sibling branches' states directly, reducing duplicated memory and exchanges of text summaries. Its trainer processes episode branches together and assigns rewards based on each branch's contribution. Trained on Qwen3.5-4B with a sparse BM25 retrieval corpus, the system also reached 51.2% on OpenResearcher. The authors report 3.3 times higher training throughput and a 2.8-fold reduction in peak memory for stored attention states.