Skills vs MCP vs RAG vs Memory: a 4-part framework for agent knowledge
MaryamMiradi · x · 2026-10-05
The author proposes separating agent knowledge into four mechanisms: Skills = procedure (repeatable task steps and judgment, loaded on demand), MCP = access (connecting to external systems and tools), RAG = knowledge, and Memory = experience.
Using a production checkout returning a 500 error as the example, the common approach—dumping runbooks, dashboards, docs, and customer history into context—fails because more context does not automatically make a better agent. The point is matching the right mechanism to the right knowledge type: Skills answer "what should I do" (check error rate → inspect recent deployments → escalate past a threshold), MCP answers "what can I reach".
Core takeaway: route different knowledge types to different mechanisms instead of stuffing everything into context.
More from coding & agent
- Opus 5.5 as daily driver, GLM for cheap prod: a practical model split — ivan_bezdomny · 2026-10-05
- Astra agent reverse-engineers macOS binaries to clone all 27 Liquid Glass effects in GPUI — steipete · 2026-10-05
- Jeff Dean's 1-hour AI engineering lecture: from LLM basics to graph-orchestrated agent swarms — glenbeer · 2026-10-05
- Fully local parkour sim vibe-coded with GLM 5.3 Flash on 2x DGX Sparks, recipe open-sourced — -dysangel- · 2026-10-05
- Someone mapped an entire company's workflows onto OpenAI Dots — a playbook solo builders can steal — Roger_M_Taylor · 2026-10-05
- Open-source Microsoft 365 MCP server brings Outlook, Teams and OneDrive control to AI agents — modelcontextprotocol · 2026-10-05