The next frontier of AI isn’t just a smarter chatbot; it’s the Foundation Agent, a system designed to transform experience into knowledge.
Drawing from the latest research in the book Advances and Challenges in Foundation Agents (Liu et al., 2025), we can now map how these agents learn by comparing them to the most sophisticated learning machine we know: the human brain.
1. The Biological Blueprint: How We Learn
The human brain doesn’t have a “single” learning button. Instead, it coordinates learning across four distinct neural systems, each with a specific job:
- The Hippocampus (Episodic Memory): Handles the rapid encoding of life experiences (the “What happened?”).
- The Cerebellum (Supervised Learning): Refines precise motor skills and coordination.
- The Basal Ganglia (Reinforcement Learning): Processes rewards and dopamine signals to tell us what to do again.
- The Cerebral Cortex (Unsupervised Learning): Extracts long-term patterns and broad understanding from the world.
2. The Digital Mirror: How AI Agents Learn
Even though AI is built on code and silicon, the way it processes information is surprisingly similar to the way we do. It uses “Mental State” updates to bridge the gap between being a blank slate and actually being useful.
The “World Knowledge” Phase (Pre-training): Think of this as the AI’s childhood. Just like a kid absorbs the world by watching and listening before they’re ever given a specific job, the AI goes through “Unsupervised Pattern Extraction.” It sifts through massive amounts of data to learn how language works and how the world is put together. It’s not trying to solve a specific problem yet; it’s just building a massive internal map of “stuff.”
The “Schooling” Phase (Fine-Tuning) Once the AI has a general grasp of the world, it goes to “trade school.” This is what the experts call Supervised Learning. We give it specific examples and feedback to sharpen its skills in a particular area like coding, medical advice, or customer service. This process actually updates the model’s “brain” parameters, making those new skills a permanent part of who it is.
The “Working Memory” Phase (In-Context Learning) This is the most human part of the whole process. Sometimes, you don’t need a permanent brain change; you just need to remember what someone told you five seconds ago. This is the AI’s version of working memory. It adapts to a task “on the fly” based only on what’s happening in the current conversation. It’s like following a recipe. You don’t need to memorize it forever to get the dinner on the table right now.
3. Mapping the Mental State (Representative Methods)
Different AI systems focus on updating different “components” of their internal mind. Research highlights several key players:
- Voyager & Generative Agents: Primarily focus on Memory Updates, allowing the agent to remember past interactions and apply them to future ones.
- Reward Agent: Focuses on Reward Modeling, teaching the agent what a “good” outcome looks like (Basal Ganglia style).
- WebDreamer: Focuses on World Model Construction, helping the agent understand the rules and physics of its environment.
The Bottom Line: Efficiency vs. Scale
The comparison reveals a fascinating gap: Human learning is incredibly efficient and deeply integrated with emotions and social context. LLM learning, while less efficient, is unrivaled in its ability to process massive datasets and synthesize formal knowledge across every domain at once.
The goal for 2026? Building agents that combine the best of both: the scale of a machine and the contextual “wisdom” of a human.
