Skills vs MCP vs RAG vs Memory

People building AI agents often treat Skills vs MCP vs RAG vs Memory as interchangeable labels, but they're not. Each covers a different job in an agent stack: reusable workflows, live tool access, document retrieval, and durable state. Picking the wrong one wastes weeks.

Skills

Skills are reusable recipes: named multi-step playbooks an agent follows when a task matches a pattern. "Triage a support ticket" or "cut a release" usually means fixed steps, the tools you expect, plus a clear done state. In agent systems they keep behaviour consistent without rewriting the prompt every turn. The agent loads the skill, runs the steps, and reports against the success criteria you defined.

They shine for repeatable ops and support flows where the path is known but the details change. A skill might call calendar tools, open a ticket, then ping Slack once checks pass. You get auditability: you can read the playbook and see what the agent is supposed to do.

But a skill is only as good as the steps you encode. When the world drifts (a new approval gate, a renamed API), brittle playbooks fail quietly unless you version and test them like any other runbook.

MCP

Model Context Protocol (MCP) is Anthropic's open protocol for connecting models to tools and data sources. MCP servers expose live connectors: databases, calendars, ticketing systems, file stores. The model discovers tools, calls them with structured arguments, and gets real results back. Adoption has spread well beyond the original stack; many agent frameworks now speak MCP as a standard way to plug in external systems.

In practice MCP is how an assistant stops being a chat box and starts acting. Need to query inventory, create a Jira issue, or read a repo? Point at an MCP server that owns that surface. You keep credentials and rate limits on the server side, not in the prompt.

I've seen teams bolt MCP on with no skills and no memory, then wonder why the agent hammers tools at random. The protocol gives you the pipe. It does not decide when to call a tool, which skill to run, or what to remember afterwards.

RAG

Retrieval-augmented generation (RAG) fetches relevant docs or chunks into context at query time, then generates an answer grounded in what was retrieved. You index a corpus (policies, runbooks, product docs), retrieve matches for the question, and pass those passages to the model. It's the standard pattern when answers must track your written knowledge rather than the model's training cut-off.

Agents use RAG mid-task: "what's our refund policy?" pulls the policy chunk; "how do we restart service X?" pulls the runbook. Done well, citations are possible and hallucinations drop because the model is constrained by retrieved text.

Retrieval quality and freshness still bound the answer. Weak chunking, stale indexes, or bad ranking produce confident nonsense that looks cited. And RAG is not memory of the user; it's search over a corpus at ask time.

Memory

Memory is durable personal or agent state that persists across turns and sessions: preferences, facts, past decisions, logs. Unlike the context window, it survives when the chat ends. An agent might store that you favour UTC, or that last week's deploy failed on billing.

That state lets the next run start smarter. Support agents skip re-asking known facts; personal assistants stop resetting your tone preferences every Monday. Memory can be structured (key-value preferences) or more free-form (summarised episode logs), but the point is persistence you control.

Without governance it becomes a junk drawer. Outdated facts linger, sensitive data accumulates, and agents act on stale preferences. You need write rules, expiry, and a way to correct mistakes.

Quick Comparison

• Term: Skills. What it is: Named multi-step playbooks / reusable agent recipes. Primary strength: Consistent repeatable workflows. Common uses: Ops runbooks, support triage, release flows.
• Term: MCP. What it is: Model Context Protocol: tool and data connectors. Primary strength: Live access to external systems. Common uses: Databases, tickets, calendars, APIs.
• Term: RAG. What it is: Retrieval-augmented generation. Primary strength: Grounded answers from your corpus. Common uses: Policies, docs, runbooks at query time.
• Term: Memory. What it is: Durable personal/agent state across sessions. Primary strength: Persistence beyond the context window. Common uses: Preferences, facts, episode logs.

How they work together

A working agent often combines all four pieces of Skills vs MCP vs RAG vs Memory. A skill defines the playbook. MCP supplies the live tools the skill needs. RAG pulls policy or runbook text mid-step. Memory stores what you learned so the next run does not start from zero. Treat them as layers, not rivals: Skills for how, MCP for what you can touch, RAG for what the docs say, Memory for what we already know about this case.

Which one do you need?

Wire live systems first? Start with MCP. Answers must cite your docs? Build RAG. The agent repeats the same multi-step job? Write Skills. It should know you next week? Add Memory. Most production assistants need more than one; choose by the gap you're closing, not the label on the slide.

Skills say how to act; MCP says what you can touch; RAG says what the docs know; Memory says what you remember. Get those four straight and the architecture stops arguing with itself.