Articles/Practice

Practice

Skills, AGENTS.md and Memory: Which Layer Should Hold What?

Agent Skills became an open standard, AGENTS.md moved to the Linux Foundation, and Claude Code now reads it. A practical map of what belongs in rules, skills and memory.

Jose R. Gonzalez, CPA

Jose R. Gonzalez, CPA · Founder of ContxOS

Updated · 2 min read

In short

Agents now have three places to keep knowledge, and they do different jobs. Rules files (CLAUDE.md, AGENTS.md) are loaded every session and hold short, always-on instructions. Skills (the open SKILL.md standard) package procedures the model loads when a task calls for them. Memory holds what the project learned: decisions, history and unfinished work. Anthropic’s AI-native SDLC playbook adds a fourth, stage artifacts such as intent.md and spec.md. Most knowledge problems in 2026 come from putting one layer’s content into another.

What happened this year

  • Agent Skills, a folder with a SKILL.md file loaded in stages, was published as an open standard in December 2025. Adopters now include ChatGPT and Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Letta. The Skills API left beta on the Claude Platform in August 2026.
  • In December 2025 MCP and OpenAI’s AGENTS.md became founding projects of the Linux Foundation’s Agentic AI Foundation.
  • Claude Code began reading AGENTS.md natively in September 2026 (by default only when no CLAUDE.md exists). Its docs are explicit that these files are "context, not enforced configuration".
  • Anthropic’s AI-native SDLC playbook (August 2026) chains stage artifacts, intent.md, spec.md and plan.md, so that "every stage commits an artifact the next stage can read."

A working map

  • Rules file: short and always true, such as commands, conventions and hard constraints. Everything here costs context on every turn.
  • Skills: how to do a kind of task, loaded when relevant. Versioned like code.
  • Stage artifacts: the intent, spec and plan for one piece of work, handed from stage to stage.
  • Memory: what the project has decided and learned over time, and where unfinished work stands. It is searched, never loaded whole.
Rule of thumb: if it is needed on every turn, put it in the rules file. If it is a how-to, make it a skill. If it is a decision or the state of the work, keep it in memory.

Where it goes wrong

The usual failure is a rules file that has become a history book, full of old exceptions, reversed decisions and notes about last month’s incident. It is loaded every turn and stays true for no one. The opposite failure is a procedure that lives only in memory, retrieved in fragments when it should be followed step by step.

What it means in practice

ContxOS sits in the memory layer and works with the others. It installs its own commands as skills, it leaves CLAUDE.md to hold rules, and it keeps procedures as workflows whose steps are read one at a time as they are executed, which is where a long procedure differs from a skill.

Sources

  1. Agent Skills — agentskills.io
  2. MCP joins the Agentic AI Foundation — Model Context Protocol blog, December 9, 2025
  3. Linux Foundation announces the Agentic AI Foundation — Linux Foundation, December 9, 2025
  4. Claude Code changelog — Anthropic on GitHub
  5. Manage Claude’s memory — Claude Code docs
  6. The AI-native SDLC playbook — Anthropic, August 21, 2026
  7. Claude Platform release notes — Anthropic

Frequently asked questions

Can a skill replace memory?

No. A skill describes how to do something. Memory records what was done, decided and learned. An agent needs both.

Give your agent a memory that lasts

Connect ContxOS to Claude in about a minute. The free plan includes one project.

Start free Setup guide

Product names belong to their owners. ContxOS is not affiliated with the companies named here.