Meta AI agent architecture developments

Meta researchers have introduced new architectures to enhance the performance and autonomy of AI agents during complex tasks.

In late September, researchers introduced the Proactive Memory Agent (PMA), a system designed to combat ‘behavioral state decay’—a phenomenon where agents lose track of subgoals during long-duration tasks. The PMA functions as a separate entity that monitors activity and selectively injects reminders into an agent’s context to prevent errors or repetition. Testing showed improved success rates on benchmarks like Terminal-Bench 2.0 and τ²-Bench when paired with models such as Claude Sonnet 4.5.

By October, Meta Superintelligence Labs developed Agentic Meta-Reasoning, an inference architecture focused on computational resource management. This system allows agents to autonomously evaluate their own progress and decide whether to continue, correct, or stop a strategy. By separating task execution from operational decision-making, the architecture aims to improve metacognitive control, with experimental tests using GPT-5.5 showing a 71.5% success rate on hidden tests.