# Skills, MCP, RAG, and Memory: The Four Ways AI Agents Actually Learn Things
Why "just add more context" isn't a strategy — and what these four building blocks are really for Every language model, on its own, only knows what was in its training data. It has no access to your company's HR system, no memory of yesterday's debugging session, and no idea how your team likes a 500 error triaged. Four mechanisms have emerged as the standard ways engineers close that gap —…
Four mechanisms have emerged as the standard ways engineers enable AI agents to learn — Retrieval-Augmented Generation (RAG), the Model Context Protocol (MCP), Agent Skills, and agent memory. Each mechanism addresses a different aspect of an agent's knowledge and capabilities.
RAG allows an AI model to expand its knowledge base by retrieving relevant information from external documents. It injects these external passages into the model's generation process alongside the original query, resulting in more specific, diverse, and factual answers. RAG is particularly useful for answering questions based on static, already-written knowledge like policy documents, runbooks, and manuals.
MCP, on the other hand, enables agents to interact with external systems and retrieve live data or even take actions. It provides a standardized protocol for connecting AI agents to databases, SaaS tools, and internal APIs. MCP allows agents to access dynamic information that is not stored in predefined documents, such as current vacation days in an HR system.
Agent Skills focus on procedural knowledge and enabling agents to perform specific tasks repeatedly. A skill is a set of instructions, often stored in a SKILL.md file, that guides an agent on how to carry out a particular task, such as triaging production errors or building slide decks. Skills use a progressive disclosure approach, where only the necessary information is loaded into the agent's context window when it is needed, preventing the context window from becoming too large.
Memory, the fourth mechanism, differs from the others as it represents knowledge that the agent acquires on its own during interactions. Unlike the curated knowledge in RAG documents or MCP systems, agent memory is built from the agent's own experiences. The most cited framework for understanding agent memory is the Cognitive Architectures for Language Agents (CoALA) model from Princeton, which divides memory into working memory (live context window) and long-term memory (semantic memory, procedural memory, and other categories).
Working memory is volatile and disappears when the session ends, while semantic memory consists of general facts and conventions. Procedural memory, on the other hand, contains how to perform specific tasks.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
