Mem-I was proposed in September 2025 by a research team from Anuttacon, the University of California, San Diego, and Stanford University. The related research findings were published in a paper. Mem-α: Learning Memory Construction via Reinforcement Learning .

Mem-I is a reinforcement learning framework that trains agents to effectively manage complex memory systems through interaction and feedback. Unlike existing methods, this framework enables agents to learn memory-building strategies for complex, multi-component memory architectures. First, researchers formulate the memory-building process as a sequential decision-making problem, where the agent processes blocks of information, decides which memory operations to perform, and receives multiple rewards based on downstream question-and-answer accuracy throughout the interaction history. Second, a specialized training dataset covering various multi-turn interaction patterns, including dialogue, document sharing, pattern recognition, and storytelling, is constructed, coupled with comprehensive evaluation questions requiring full memory to answer correctly.