CLAUDE.md and rules files vs a shared memory.
An instructions file tells an AI coding tool how one repository works. A shared memory holds what your team decided everywhere else.
The short answerKeep CLAUDE.md, AGENTS.md, Cursor rules and Copilot instructions for what is true of one repository: how to build it, test it and lay it out. Use a shared memory like Contexel for what is true across repositories, apps and people: team conventions, decisions and context that was never written in the code.
At a glance
| Instructions files | Contexel | |
|---|---|---|
| Examples | CLAUDE.md, AGENTS.md, .cursor/rules, .github/copilot-instructions.md, GEMINI.md | One memory every connected app reads |
| Scope | One repository, or one person's machine | Every repository, every app and every person you share a workspace with |
| How the AI reads it | The whole file, at the start of every session | Only the parts relevant to the question, when it needs them |
| Who updates it | Whoever edits the file and gets the change merged | Anyone in the workspace, from any connected app, with the change history kept |
| Outside coding tools | Not read by ChatGPT or Claude on the web | Read by ChatGPT and Claude too |
What instructions files do well
An instructions file lives next to the code, is reviewed like code, and loads every time. For "run pnpm test, never npm" or "components go in src/ui", nothing beats it. Keep using it.
Where files run out
- They grow.Every rule is loaded into every session, relevant or not. Past a few hundred lines, the whole file rides along with every prompt and the important lines get lost.
- They get copied.A convention that applies to ten repositories ends up in ten files, and they drift apart.
- They miss decisions made elsewhere.What was agreed in a chat with Claude, in a meeting, or by a teammate in a different editor never reaches the file unless someone copies it in.
- Each tool reads its own file.Claude Code, Cursor, Copilot and Gemini CLI each look for a different file name.
What a shared memory adds
Contexel holds the knowledge that belongs to the team rather than the repository, and hands each AI only the part a task needs. On our own test of a sample team's memory (39 tasks, averaged), Contexel sent the AI about a third as much text as pasting the whole instructions file: about 1,400 tokens against 4,200. A token is a few letters of text, the unit AI models read and are billed by. Results depend on your data.
You can start from what you already wrote down: paste a CLAUDE.md or AGENTS.md into the console, and Contexel splits it into one memory per section. Then trim the file back to what is truly about that repository.
Using both
The setup that works: a short instructions file per repository, plus one line in it telling the AI to check Contexel for team decisions. The console writes that line for you on its Connect tab, for Claude Code, Cursor, GitHub Copilot and Gemini CLI.
Questions
Should I delete my CLAUDE.md?
No. Keep it for instructions about one repository. Use Contexel for what your team knows across repositories and apps.
Does my code get uploaded?
No. Contexel only stores what you or your AI choose to save as a memory. It does not read your repository.
Can I import an existing instructions file?
Yes. Paste a CLAUDE.md, AGENTS.md or any notes file into the console, and Contexel splits it into separate memories, one per section.