Meta Gives AI Agents a Second Brain: Persistent Memory

Meta has taught AI agents to stay on track through long, complex tasks. In the new system, one agent does the work while a second — the memory agent — monitors execution and sends targeted hints at the right moment: what's already been tried, what errors have been found, and what should not be touched. On the Terminal-Bench 2.0 benchmark, task completion rates rose from 38% to 46%; for service tasks, from 55% to 62%.

Memory loss is one of the main reasons agents stall and keep repeating the same mistakes. Meta has released the solution as a plug-and-play module that connects to any agent without retraining — and that raises the bar for real-world autonomy.