A separate agent that remembers what matters can make AI agents noticeably better at long tasks. Researchers at Meta, led by first author Yifan Wu, describe the approach in a paper titled "Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents."
The problem they target is familiar. When an agent works on one task for a long time, key details from the start get buried under new steps or fall out of context. The authors call this behavioral state decay.
How it works
The memory agent keeps a structured memory bank about the task: facts, details of the environment, fixes that worked, commands that failed, and what is still unfinished. It updates the bank from the most recent steps. At each step, it decides whether to hand the main agent a short reminder or stay silent. The main agent itself is not modified.
The team tested Claude Sonnet 4.5, Claude Opus 4.6, and Qwen3.5-122B as main agents. The memory agent was either Claude Opus 4.6 or Qwen3.5-27B, further trained on command-line tasks.
Results
On Terminal-Bench 2.0, a set of 85 command-line tasks, Claude Sonnet 4.5 rose from 37.6 to 45.9 percent solved with reminders. Opus 4.6 gained less, from 43.5 to 45.9 percent.
On τ2-Bench, 278 customer-support tasks for airlines, retail, and telecom, Sonnet 4.5 climbed from 55 to 61.8 percent and Opus 4.6 from 66.2 to 68.7 percent. The abstract reports peak gains of 8.3 points on Terminal-Bench and 6.8 points on τ2-Bench.
According to the paper, targeted reminders beat both constantly injecting the full memory and plain retrieval from the notes.
