# Procedural Memory Is Not All You Need: Bridging Cognitive Gaps in LLM-Based Agents

**Authors:** Schaun Wheeler, Olivier Jeunen  
**arXiv:** [2505.03434](https://arxiv.org/abs/2505.03434)  
**Date:** 2025-05-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.03434v1/

## Key points

- Large Language Models (LLMs) face constraints in navigating intricate and unpredictable environments due to their reliance on procedural memory.
- To excel in "wicked" learning environments with shifting rules and ambiguous feedback, LLMs need to be enhanced with semantic memory and associative learning systems.
- Decoupling cognitive functions and adopting a modular architecture can bridge the gap between narrow procedural expertise and adaptive intelligence for real-world problem-solving.
- Designing application and data systems with the learning environment in mind is crucial, as different environments may require different capabilities.
- Kind environments may only necessitate procedural capabilities, while wicked environments demand hybrid architectures integrating associative and semantic reasoning.
- Investment in research into neural-symbolic architectures, sparse memory models, and other frameworks that facilitate explicit reasoning and adaptive learning is advocated.
- Developing agents that complement human ingenuity in uncertainty rather than replicating procedural expertise is key.
- Shifting from monolithic to modular architectures represents a necessary evolution in AI design to address the complexity of real-world decision-making.

## Summary

The paper delves into the constraints of Large Language Models (LLMs) when navigating intricate and unpredictable environments due to their reliance on procedural memory. It argues that to enable agents to excel in such "wicked" learning environments, where rules are constantly shifting and feedback is ambiguous, LLMs need to be enhanced with semantic memory and associative learning systems. By decoupling these cognitive functions and adopting a modular architecture, we can bridge the gap between narrow procedural expertise and the adaptive intelligence required for real-world problem-solving. The authors emphasize the importance of designing application and data systems with the learning environment in mind, highlighting that different environments may require different capabilities. While kind environments may only necessitate procedural capabilities, wicked environments demand hybrid architectures that integrate associative and semantic reasoning. Furthermore, the paper advocates for investment in research into neural-symbolic architectures, sparse memory models, and other frameworks that facilitate explicit reasoning and adaptive learning. By embracing these principles, we can develop agents that complement human ingenuity in uncertainty rather than merely replicating procedural expertise in structured domains. This shift from monolithic to modular architectures represents a necessary evolution in AI design, acknowledging the complexity of real-world decision-making and addressing the limitations of current approaches. The goal is to create intelligent agents capable of thriving in dynamic and challenging environments by combining procedural expertise with semantic understanding and associative reasoning.