Which tool to choose
All the main assistants offer these three layers, under different names. On ChatGPT you have custom instructions, memory and Projects. On Claude, Projects have a knowledge base where you upload files, plus a memory in the form of readable notes. On Gemini, Gems are tailored assistants with stable instructions, and they can connect to your documents on Drive, updating themselves when you modify them. Choose the one you already use: the way to combine the three layers is similar everywhere.
How it's done
- Set the custom instructions once and for all: who you are, what you do, how you want the answers. It's the baseline behavior, valid in every chat.
- Turn on memory and tell it the recurring facts about you ("remember that...").
- For every serious job, create a dedicated project or space, with its own specific instructions and reference files (brief, documents, examples).
- Work inside that project: every conversation already starts with instructions, memory and files in view, without starting from scratch.
A concrete example
Sara, a freelancer, worked for three different clients and each time re-explained tone, sector and constraints to the AI. She built a layered persistent memory: in the custom instructions she put who she is and how she writes in general; then for each client she created a project, with the brief, the examples of approved texts and the specific rules inside it. Now, when she has to write for "Client X", she opens the right project and the AI is already perfectly immersed in that context. No more preambles: she writes straight away in the way needed.
When it does NOT work (and how to fix it)
If the contexts of different jobs get mixed up
Putting everything in a single global memory makes the AI confuse one client with another, one project with another. The fix is to separate: the general instructions in the base memory, but the specific context of each job inside its dedicated project, isolated from the others.
If the context to remember is too much for memory
Memory is made for brief facts, not for whole documents. If you have to keep pages of material in mind, don't cram them into memory: upload them as files into the project space, which is designed exactly for that.
If you change assistant
The memory and projects of one tool don't pass automatically to another. If you anticipate possibly switching, don't rely solely on the internal memory: keep the important context in your own documents too, which you can reload anywhere.
A tip from someone who really uses it
The most solid persistent memory doesn't live inside an AI, it lives in a context document that you own: a file with who you are, the rules, the state of the project. You load that document into any project and any tool, and in an instant you bring the AI up to speed. That way you're not held hostage by the way each assistant handles memory, and if you change tool you don't start from scratch.
Frequently asked questions
Is giving memory to the AI like training it on me?
No. The model doesn't change and doesn't learn your person: it stays identical for everyone. What you build is supporting material — instructions, notes, files — that the AI rereads at every start. It's a reference memory, not learning.
How much context does a project hold?
Much more than a single chat, but it isn't infinite: every project space has a ceiling on how much material it can keep in view. For normal jobs it's amply sufficient; for huge archives, it's better to select the most relevant documents instead of uploading everything.
Is it safe to keep information in memory and in projects?
It's data on your account, which you control, but it's on a server: the usual caution applies. Keep out passwords, very sensitive data and third-party information, and for confidential work contexts consider plans with data guarantees.
Does persistent memory mean the AI learns and improves on me over time?
No, and it's the misunderstanding to clear up. The AI doesn't get better at you: it accumulates notes and files that it rereads, but the competence stays the same. Memory improves over time only if you're the one tending it — updating the context document, correcting the notes, keeping the projects in order. It's not it that grows by getting to know you: it's you who build and maintain what serves as its memory.