Projects and memory

A new chat starts empty. Everything the assistant nevertheless knows about you comes from one of three levels – and most everyday annoyances arise because something sits on the wrong one.

Level 1: standing instructions

In the settings of every major assistant there is a field for permanent instructions. What is written there applies in every conversation. So only what is truly always the case belongs in it: your role, the language, two or three style rules.

I run an online store for bicycle parts with four employees.
Answer in English, informally, without introductory filler.
If a detail is missing, ask instead of guessing.

Three lines. Anything that applies to one subject only is wrong here – otherwise it colours your holiday planning too.

Level 2: projects

A project is a folder with its own context: its own instructions, its own files, its own conversations. The assistant reads the project’s documents along with every question without you uploading them again.

A good project has three ingredients:

1. Purpose in one sentence. “Product copy for the store is written here.”
2. Rules that apply only here. Tone, length, banned words.
3. Two or three examples showing what a good result looks like. Examples work harder than any description.

Create a project as soon as you paste the same context for the third time. And keep it lean: twenty half-fitting files make answers worse because the assistant does not know which one holds. Why that is so is explained in Context engineering.

Level 3: memory

On top of that, assistants note things from your conversations by themselves or search earlier chats. That is convenient – and the level you control least as long as you do not look.

Three moves everyone should know:

• View. The settings show what is stored. Read through it once a quarter.
• Delete and correct. Entries can be removed one by one. A sentence like “Forget that I …” in the chat often does the job too.
• Switch off for one chat. For subjects that should not stick there is a temporary mode without memory.

What belongs where

The last row is meant seriously. Health, finances, other people’s data: what is not stored cannot later surface in an answer where it has no business. The rules for that are in ChatGPT and privacy.

The limit of these three levels

All of this lives with the provider. Switch assistants and it stays behind. If you want to build knowledge that belongs to you and that any model can read, you need your own files – that is the learning path Knowledge & context.