AI agents keep forgetting their own homework between sessions - a new runtime aims to fix that.
Researchers built a runtime called AstronOS that gives long-running AI agent tasks a persistent, versioned state instead of relying on a single conversation thread. They tested five ways of handing a multi-stage software-update plan to a fresh model session: rereading the original materials, replaying the full conversation history, a deterministic text summary, a deterministic JSON summary, and AstronOS's own runtime-mediated handoff. Across 150 test runs covering ten tasks with three repetitions each, every method did fine on simple single-stage tasks. On the harder three-stage batch, though, AstronOS passed 14 of 15 runs, rereading the source material passed zero, replaying full history passed just 2 of 15, and both summary approaches failed all 15.
That gap matters because rereading or summarizing context is the default workaround most agent builders reach for today, and it looks fine on simple benchmarks. This paper suggests that shortcut quietly falls apart once a task spans multiple stages. AstronOS also used fewer tokens per successful run than the other methods, though it took longer per attempt, pointing to structured state tracking rather than bigger context windows as the more durable fix.
It is a narrow benchmark built and graded by the team that made the runtime, so 14 out of 15 is a promising number, not proof the amnesia problem is solved.