Meta Unveils AI Agents That Combine Memory and Continuous Reasoning

According to AI TIMES, Meta on October 12 unveiled next-generation AI agent models, "Muse Wonder" and "Muse Wonder Full," which combine long-term memory and continuous reasoning, extending AI for consumers and businesses in a way that lets it complete complex tasks from start to finish. The models are designed to handle up to 1 million tokens of context and to continue tasks such as document review, research, and coding by using browsers and multiple applications. The key features are memory, which pulls in and uses needed information across multiple conversations, and task persistence, which keeps a goal moving forward by finding alternative paths even when errors or variables arise midway. Meta said it advanced pretraining and reinforcement learning for this effort and placed greater emphasis on managing execution flow than on generating answers. It tested performance in specialized fields such as law, finance, health care, and software development, but said some evaluations included its own results, leaving room for interpretation. The rollout will also begin with developers, with initial availability through APIs and code tools, and whether research-stage performance can be reproduced reliably in real-world work environments has emerged as the next point to watch.