Meta's PAMO trains personal agents to use cross-app history in recommendations

Meta researchers introduce PAMO, a training method that teaches a personal AI agent when to use a user's history from other apps to improve an app's recommendations.

In tests on real Steam users with Nintendo and Xbox history, a 4B PAMO agent reached a 47.9% top-10 hit rate, compared with 40.8% for the platform's own ranking, according to a Meta paper reported by Rohan Paul on X. The agent also outperformed standard reinforcement learning training on all five test sets.

To measure reliance on cross-app history, the researchers hid that history and checked which of the agent's answers changed, according to the same report.