Definition
Plain language
A system that lets an AI research agent build lasting understanding from its experiments instead of forgetting its own lessons over a long run.
As stated in the literature
An autonomous-research framework structured around a hypothesis tree with insight propagation, isolated single-hypothesis executor agents, and a held-out merge gate; reported to beat Codex and Claude Code on six research-optimization tasks at comparable budget.
Why it matters: It lets a long-running research agent accumulate and reuse what it learns, so a multi-step investigation gets smarter rather than repeating itself.
For example, when one experiment reveals that a smaller learning rate helps, Arbor records that lesson so later experiments in the same run can build on it instead of rediscovering it.
Heard on the show
“… joint team at Renmin University of China and Microsoft Research, and the system they built is named Arbor. …”Episode 131 — Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix