SciMARS Memory Lifts Multi-Agent Science Benchmark Scores
Paper reports gains from shared memory in multi-agent science system
Researchers describe SciMARS: Multi-Agent Recursive Self-Improvement for Closed-Loop Scientific Discovery in a paper dated November 8, 2026, with an update posted November 10, 2026, on arXiv (arXiv). The work tests whether giving a multi-agent system a shared memory module improves its ability to complete scientific-discovery and software-agent tasks compared with a base multi-agent team running the same benchmarks.
Benchmark results
On ScienceAgentBench, the SciMARS + Memory configuration solved 14 of 38 tasks, versus 11 of 38 for the base multi-agent team, according to the paper (arXiv). On AppWorld, a benchmark of agentic software tasks, SciMARS + Memory solved 11 of 20 tasks compared with 5 of 20 for the base team — more than doubling the success rate.
On the paper's own Autonomous Science Bench, results varied sharply by underlying model. Using GPT-6-Astra, SciMARS + Memory solved 12 of 18 attempts versus 8 of 18 for the base team. Using GLM-5.3, the memory-equipped system solved 2 of 18 attempts versus 1 of 18 for the base team — a smaller improvement that still favored the memory configuration (arXiv).
Caveat: no new wet-lab experiments
The evaluation did not involve newly executed physical laboratory experiments. Instead, the authors used experimental-data replay and a digital twin to simulate scientific discovery tasks, a methodological detail flagged in coverage of the paper's claims (rsi.watch, https://rsi.watch/self-improving-ai/). That distinction matters for how the reported gains should be read: the benchmarks measure an agent team's ability to navigate scientific workflows and reproduce or extend results from existing data, not its ability to run original experiments in a physical lab.
Why it matters for the agent economy
Recursive self-improvement and shared memory are recurring themes in recent agent research, with related work on self-improving agents for molecular discovery and unified frameworks for agent-guided science appearing around the same period (themoonlight.io, themoonlight.io). SciMARS adds to a growing body of benchmark evidence that persistent memory across agent iterations can meaningfully raise task-completion rates on structured, multi-step benchmarks — even as the field has yet to validate these systems against real, unreplayed wet-lab science. The available sources do not name the paper's full author list or provide an official institutional statement on the findings.


