Large language models (LLMs) have shown incredible potential, but they aren't enough to achieve Artificial General Intelligence (AGI) on their own. A major roadblock? Their limited effective context length, even with claims of handling millions of tokens. Simply stuffing an LLM with endless raw data won't work. It turns out that finding a needle in a haystack of information, while also conducting complex reasoning, is nearly impossible for today's LLMs. Researchers are now exploring the idea of 'AI-native memory' as the missing puzzle piece. Imagine AGI as a computer system: the LLM is the processor, its context is the RAM, but we're missing a hard drive - the long-term memory. This 'memory' wouldn't just store raw data like current retrieval-augmented systems, but would also house inferences derived from that data, making the LLM's job far easier. In the near term, this memory could take the form of natural language descriptions, almost like building a 'Memory Palace.' Ultimately, though, researchers envision each user having a personalized, AI-native Lifelong Personal Model (LPM). This LPM would be a neural network, possibly even an LLM itself, constantly learning and compressing all of a user's memories, even those beyond language. This personalized memory could revolutionize how we interact with AI, enabling truly personalized experiences, recommendations, and even predicting user behavior. The challenges lie in training and serving these personalized LPMs efficiently. Researchers are investigating techniques like LoRA for parameter-efficient fine-tuning and developing new serving frameworks for cost-effectiveness. There's also the 'cold start' problem for new users and the need to address catastrophic forgetting as the LPM learns. The first steps towards L2 LPMs are already being taken. Researchers are using advanced LLMs like GPT-4o to synthesize training data and developing methods to ensure LPMs prioritize individual user needs. The potential of AI-native memory is vast. It represents a fundamental shift in how we think about LLMs, moving from isolated models to integrated systems with personalized memory banks. This could be the transformative infrastructure that unlocks the next level of AI, powering personalized interactions, proactive engagement, and even social connections in the AGI era.