Long‑horizon agents routinely lose track of task requirements, environment facts, and subgoals, a failure mode the authors name behavioral state decay: “We call this failure mode behavioral state decay: during long-horizon execution, information that should shape future actions like task requirements, environment facts, previous attempts, failure diagnoses, intermediate discoveries, and open subgoals stops influencing the agent’s next decision.” [1] A tiny auxiliary memory agent that injects reminders only when needed can halt that drift.
Previous long‑horizon systems rely on passive context expansion or static retrieval to keep information accessible, treating memory as a read‑only store rather than an active decision partner. Those approaches assume the transformer’s context window will eventually contain everything needed, which quickly breaks down as trajectories grow beyond its limits.
Memory intervention lifts pass@1 by 8.3 percentage points on Terminal‑Bench 2.0 for the strong Sonnet 4.5 action model: “Memory intervention improves pass@1 on both benchmarks and across both action-agent strengths: on Terminal-Bench 2.0, Sonnet 4.5 gains +8.3 pp () …” [1] The gain persists across a range of task complexities, showing that selective reminders can recover information that would otherwise be lost.
On the τ²‑Bench suite the same module adds 6.8 percentage points to Sonnet 4.5’s task‑weighted average pass@1: “on -Bench, Sonnet 4.5 gains +6.8 pp on the task-weighted average () …” [1] This improvement mirrors the Terminal‑Bench result, suggesting that proactive memory benefits are not confined to a single benchmark family.
A SETA‑trained proactive memory policy raises a frozen action agent’s pass@1 on held‑out Terminal‑Bench 2.0 from 37.6 % to 41.1 %: “On held-out Terminal-Bench 2.0, the SETA-trained memory agent improves the frozen action agent from 37.6% to 41.1% pass@1.” [1] Even without fine‑tuning the action model itself, an open‑weight memory learner can extract measurable performance gains.
The study does not explore how the memory bank scales with longer contexts or larger model families, limiting confidence about runtime overhead and storage costs in production settings. Ablations demonstrate that selective intervention beats always‑on injection, yet they evaluate only up to 27 B parameters and modest trajectory lengths; it remains open whether the same pattern holds for multimodal agents or tasks with millions of steps.
Future long‑horizon benchmarks should include a proactive memory baseline rather than treating raw context size as the sole variable, because the evidence shows that an active reminder system can reliably close the gap caused by behavioral state decay.
Top comments (0)