Meta Builds a ‘Second Brain’ for AI Agents with Persistent Memory

Meta’s developed a method for training AI agents to carry out long and complex tasks. Their approach involves having one agent perform the task while another (the “memory agent”) monitors the first agent’s progress and gives it advice along the way, like reminding it of past experiences, errors it has committed, and what not to do. Using this technique, they were able to boost the success rate of tasks on the Terminal-Bench 2.0 benchmark from 38% to 46%, and for service tasks, from 55% to 62%.

One of the main reasons why agents fail and make repeated errors is because they forget things. Today, they’re releasing this technology as a plug-and-play module that can be added to any agent without requiring any additional training. It’s the future of autonomy.