Which tool to choose

You need project spaces: Projects on ChatGPT and Claude, Gems on Gemini, which let you keep an isolated context for each job. The global memory, the one that applies to all chats, should be reserved for the facts about you that don't change from project to project. If your tool doesn't have projects, the alternative is a separate context document for each job, which you load only when you work on that one.

How it's done

  1. Create a distinct project for each job or client, well separated from the others.
  2. In the global memory put only what is common to everything: your role, your baseline style, general preferences.
  3. The specific context of a project — brief, rules, that client's files — goes inside its project, never in the global memory.
  4. When you switch from one job to another, open the right project: that's where the context of that job lives, isolated from the rest.

A concrete example

A consultant was working with three very different companies, and at first he kept everything in scattered chats with memory active. The result: the AI advised him for one company with strategies meant for another, mixing the contexts. He redid the setup: in the global memory only "I'm a marketing consultant, I write in a concrete way"; then a project for each company, with that company's sector, data, and constraints alone inside. Since then each project answers him grounded in the right context, and the contaminations have vanished: opening the project is like walking into the right client's room.

When it does NOT work (and how to fix it)

If the AI confuses one project with another

Almost always you've put specific context into the global memory, which applies to everything and so spills over. The fix is to move it: a client's context goes only into its project, and the global memory should be kept lean, with only the common things.

If you find yourself repeating the same instructions in every project

If you rewrite a rule identically in all projects, it means it's common and belongs in the global memory or instructions, not duplicated. Keep in the projects only what distinguishes them, and at the global level what they share.

If you don't remember which project is at what point

With several jobs in parallel it's easy to lose the thread of each. Keep in each project a short status document — where we are, what's missing — so when you reopen it you realign in a few seconds.

A tip from someone who actually uses it

Think in two clear levels: the global level is what always applies, regardless of the project — who you are, how you write; the project level is what applies only to that job. Keeping these two planes distinct, without letting them spill over, is the only thing that truly prevents contamination when you have many projects open. Every time you're about to write a piece of information, ask yourself: does it apply to all my jobs or only to this one? The answer tells you where to put it.

Frequently asked questions

Does the global memory interfere with the individual projects?

It can, if you stuff specific context into it: being global, that context enters all projects and mixes them. Kept lean, with only the common facts about you, it instead helps without contaminating. The problem isn't the global memory itself, it's what you put in it.

How many projects can I keep in parallel?

Many: the practical limit is your ability to manage them, not the tool. Having one project per job is exactly the right way to use them, because each holds its own context separately without interfering with the others.

How do I keep the contexts from mixing?

Single rule: specific context only in the project it belongs to, never in the global memory. If you respect this separation, the projects stay isolated. Contamination always comes from a piece of specific information ending up at a level that's too "high."

The more context I put in the global memory, the better for all the projects, right?

No, and it's the reasoning that creates the confusion. The global memory is shared: everything you put in it enters every project, even where it has nothing to do with anything. Filling it with specific details doesn't enrich the projects, it contaminates them with each other. The global memory should be deliberately kept poor and general; the richness of context belongs in the individual projects, each on its own.