Advances & challenges in foundation agents: Section 1.3.3 – Biological inspirations

This article is Chapter 1, Section 1.3.3 of a series of articles featuring Liu and colleagues’ book Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems.

Liu and colleagues’ agent framework, though fundamentally computational, finds its roots in biological systems, especially the intricate operations of the human brain. By exploring these biological parallels, we gain deeper insight into how Liu and colleagues’ framework can embody key cognitive faculties that make biological agents so versatile, robust, and adaptive.

Memory: capturing experience and knowledge (hippocampus and neocortex). Human memory is famously rich, vivid, and multifaceted. Neuroscience describes the hippocampus as a library of episodic memories¹ (specific, detailed snapshots of past events) while the neocortex houses semantic knowledge², the enduring conceptual understanding we build up through a lifetime. Together, these brain structures orchestrate the flow from immediate experiences to long-lasting wisdom.

Inspired by this partnership, Liu and colleagues’ agent’s memory module, M_t^mem, mirrors this duality by combining short-term memory (to rapidly record ongoing interactions and context) and long-term storage (to consolidate and generalize from past experiences). Just as the brain filters and retains information based on relevance, Liu and colleagues’ memory module prioritizes which experiences merit long-term preservation, preventing clutter and ensuring essential insights remain accessible when most needed.

World model: imagining and predicting the future (predictive processing). One of the brain’s remarkable feats is its ability to constantly predict what will happen next. Cognitive theories propose that our cortical networks operate like sophisticated prediction engines³, anticipating sensory inputs and swiftly adjusting internal expectations⁴ based on mismatches. This predictive dance allows humans to navigate effortlessly through uncertainty, imagining consequences, weighing alternatives, and adjusting in real-time.

Liu and colleagues’ agent’s world model, M_t^wm, captures this predictive prowess by continually maintaining and updating an internal simulation of the environment. It anticipates future states, evaluates potential outcomes of actions, and dynamically updates itself based on fresh observations. Much like the brain seamlessly integrates perceptions into predictions, Liu and colleagues’ agent leverages a similar Bayesian-like process to ensure its internal model stays accurate, relevant, and useful.

Emotion: guiding behavior beyond pure logic (limbic system). Contrary to stereotypes, human emotions are not irrational distractions; rather, they act as powerful and nuanced guides for decision-making. Structures in the limbic system, such as the amygdala and hypothalamus, shape our focus, modulate our reactions, and enhance learning through emotional significance^5,6. For instance, fear sharpens our attention to danger, and excitement motivates exploration and creativity.

To emulate this adaptive guidance, Liu and colleagues’ agent incorporates an emotion component, M_t^emo. Although computational emotions do not equate to genuine human feelings, this module mimics their functional role: prioritizing attention, adjusting urgency, and directing learning efforts based on internal affective states. Like the limbic system subtly steering human choices, this emotional mechanism ensures Liu and colleagues’ agent remains responsive, adaptive, and aligned with its context.

Goals and reward: shaping intentions and motivations (prefrontal & subcortical circuits). Goals give purpose; rewards reinforce behavior. In humans, abstract goal-setting and sophisticated long-term planning are primarily orchestrated by the prefrontal cortex^7,8, while reinforcement signals from subcortical regions (especially dopaminergic pathways) continuously adjust our motivations and habits⁹. This integrated circuitry enables humans to sustain complex intentions and refine actions through continuous feedback loops.

Liu and colleagues’ agent mirrors this sophisticated partnership with goal (M_t^goal) and reward (M_t^rew) components. Goals shape the agent’s overarching intentions, guiding decisions towards desired outcomes, while the reward component continuously tunes behavior and reinforces effective strategies. Together, they form an adaptive motivational system that mirrors how the brain’s goal-driven deliberation is harmonized with reward-driven adaptation, enabling flexible, contextually appropriate behavior.

Reasoning, planning, and decision-making: executive control (prefrontal cortex). The hallmark of human intelligence is arguably our capacity for reasoning and planning—deliberate, future-oriented cognition primarily governed by the prefrontal cortex. This brain region^10,11 synthesizes memory, perception, emotional states, and reward signals into coherent strategies for action. It’s what lets us imagine multiple outcomes, judge their merits, and confidently choose a path forward.

Reflecting this remarkable capability, Liu and colleagues’ agent’s reasoning sub-function acts as the executive core. It orchestrates internal simulations, weighs alternative actions, and selects optimal strategies—echoing how the prefrontal cortex evaluates possibilities before making a decision. By clearly distinguishing long-term planning from moment-to-moment decision-making, Liu and colleagues’ agent can flexibly switch between reflective deliberation and swift, intuitive choices, just as humans seamlessly alternate between careful thought and rapid reaction.

These biological analogies enrich Liu and colleagues’ computational framework by grounding its functional logic in neuroscientific realism. Yet, importantly, the parallels remain flexible, not rigidly bound to biological exactitude. They serve as guideposts rather than blueprints, highlighting fundamental cognitive principles that, when embedded into artificial agents, can yield intelligent systems capable of rich, adaptive behaviors. In subsequent chapters, we delve deeper into these modules, further exploring how they interact and evolve, guided by insights from both neuroscience and cutting-edge AI research.

Next part: Section 1.3.4 – Connections to existing theories.

Article source: Liu, B., Li, X., Zhang, J., Wang, J., He, T., Hong, S., … & Wu, C. (2025). Advances and challenges in foundation agents: From brain-inspired intelligence to evolutionary, collaborative, and safe systems. arXiv preprint arXiv:2504.01990. CC BY-NC-SA 4.0.

Header image: AI is Everywhere by Ariyana Ahmad & The Bigger Picture / Better Images of AI, CC BY 4.0.

References: