Anthropic has released research that breaks down the implementation logic behind Claude's agent memory system and, inspired by the brain's sleep-based memory consolidation mechanism, proposes an offline computing paradigm called "Dreaming." During idle interaction periods, the AI reviews historical logs, corrects memory errors, and consolidates structured knowledge, thereby addressing pain points such as agent memory confusion across sessions in long-horizon tasks. This agent memory refers to experiential knowledge that the AI acquires from interactions and tasks and that can persist across sessions; it does not modify model weights and instead relies on external storage for consolidation. The "dreaming" mechanism undertakes five core tasks, including correcting memory errors and building associative knowledge representations, while also requiring strict control over the amount of preloaded memory. Multiple sets of experiments verified the effectiveness of this memory system, yielding positive results in benchmark tests, long-horizon game simulations, and large-scale enterprise document scenario tests. However, current agent memory still faces unresolved challenges, such as competing with the main task for compute resources, additional consumption of computational resources, and difficulty in accurately predicting information value in advance. The team will continue to optimize related strategies and improve the memory infrastructure for long-term autonomous agents.