Meta AI adds a second agent to help others avoid repeat mistakes

Meta AI is testing a system where a separate 'memory coach' agent maintains a structured memory bank for a main AI agent on long tasks, deciding when to remind it of past errors and when to stay silent. The system improved benchmark scores by up to 8.3 percentage points across two tests.

Why it matters: This targets a persistent weakness in long-running agents — forgetting or repeating diagnosed errors — and the measured benchmark gains suggest memory architecture, not just bigger models, may be a key unlock for reliable multi-step agent tasks.