Imagine an AI that not only learns from experience but also generates its own memories, refining them through practice and feedback, much like humans do. This intriguing concept is now a reality, thanks to a groundbreaking method called In-Context Abstraction Learning (ICAL). Traditionally, AI models like LLMs and VLMs, while capable of in-context learning, have relied on high-quality examples provided by humans. ICAL changes this paradigm by allowing AI agents to create their own improved examples from suboptimal or even flawed demonstrations. The secret lies in distilling experiences into “programs of thought.” ICAL agents analyze suboptimal examples, identifying inefficiencies, correcting mistakes, and making important observations, such as causal relationships between actions and the way objects change over time. This process is further enhanced by feedback from human observers. When an agent tries to execute a learned program in a similar situation and fails, the human provides feedback to correct its behavior. The agent then refines its program of thought accordingly. This iterative feedback loop significantly boosts performance and improves the agent’s capacity for future action planning. The results are impressive across diverse tasks. In tests of dialogue-based instruction following in virtual household environments, ICAL agents showed a remarkable 12.6% improvement in goal achievement compared to existing methods. They also excelled in complex web tasks, increasing success rates from 14.3% to 22.7%, and proved competitive with fully supervised models in anticipating actions in egocentric videos. Interestingly, as the agent’s library of self-generated examples grows, it requires less feedback and fewer tries to learn from new demonstrations, indicating greater learning efficiency and reduced dependence on human intervention. ICAL offers a compelling vision for the future of AI, one where agents can refine their knowledge autonomously and accelerate their own learning. While challenges remain, including handling severely misleading demonstrations and addressing visual grounding issues in current VLMs, ICAL presents a significant leap towards more efficient, robust, and self-reliant AI systems. This new paradigm has the potential to revolutionize many applications, from domestic robots to personalized AI assistants, by enabling AI to adapt and learn in a more human-like way.