What is AI memory?

AI memory means two different things at once — what an assistant can see inside one conversation, and what it keeps across conversations in its own service. Here's what each one covers, and what neither does.

The short answer

AI memory is anything an assistant can draw on beyond the message you just sent. In practice it means two separate things: the context window, which is what it can see inside the conversation you're having, and built-in memory, which is what it keeps across conversations inside its own service.

The two stop at different points, which is why "can it remember?" gets answered both ways. One is limited by the size of a conversation. The other is limited by the edge of a service.

The two kinds of AI memory

AI memory splits into the context window — what fits inside one conversation — and built-in memory — what carries across conversations inside one service.

The context window is how much an assistant can look at while writing a single reply. Everything it uses for that answer has to fit inside it: your messages, its own, anything you pasted. When a conversation grows past that, the oldest parts fall out of view first. That's why the opening of a long chat goes fuzzy while the last few exchanges stay sharp — what that feels like day to day is its own post.

Built-in memory is what an assistant keeps across conversations, inside its own service. ChatGPT's memory works this way — it picks things up from your chats and carries them forward. Claude's Projects take a different route: material lives in a project, and chats started there can use it. Both are real features, and both have grown steadily.

So "it forgot" usually means one of two things. Either the conversation ran past the window, or the thing you had in mind was never kept in the first place.

What AI memory doesn't cover

Neither kind crosses from one service to another. What ChatGPT keeps, Claude can't see, and the reverse holds too. Each was built for its own users, and there's no reason for one to go reading a competitor's storage. Carrying context between two assistants runs into this first.

There's a second edge, and it's easier to miss. You don't decide all of what's kept. Built-in memory picks up what looks worth holding on to. That's convenient. It's also not the same as marking something yourself and knowing it will still be there in six months.

Neither of these is a defect — they follow from where each service's boundary sits. Which also means no feature release from either side closes the gap: a store that spans two services can't live inside either one.

The part you keep yourself

What has to live outside the assistants is usually called a knowledge layer.

A knowledge layer is a store that sits outside any single assistant and holds what you chose to save, so any assistant you connect can bring it back later. Assistants come and go; the layer stays. And nothing is in it except what you put there.

Two things keep a store like that from becoming filing work.

Semantic retrieval finds things by meaning rather than by exact words. Ask about the pricing decision you made last spring and the closest match comes back, even though it was saved under a title you'd never have guessed. Getting that to work depends on what you save, and in what shape.

MCP (Model Context Protocol) is an open standard for connecting assistants to outside tools and stores. If a store speaks MCP, any assistant that supports the standard can reach it — which is what keeps a knowledge layer from belonging to one assistant.

Three things it isn't:

• Not a notes app. Nobody goes in to read it — the reader is an assistant, which is why it doesn't matter that it's untidy.
• Not automatic collection. Saving is a decision you make. Your conversations aren't being read in the background.
• Not a filing system. No folders, no tags. The point isn't to organize it for you — it's that there's nothing to organize.

When built-in memory is enough

If you use one assistant and don't expect to switch, built-in memory covers it. It reaches further than it used to, it costs nothing extra, and there's nothing to connect.

The reasons to keep something outside show up here:

• You use more than one assistant, and you've caught yourself telling the second one what you already told the first.
• You keep re-explaining the same background — the constraints a project always has, how you like things written, what you've already ruled out.
• You might switch later. Nothing about that feels pressing until the switch is actually in front of you.

If it does apply to you, the two things that actually differ between the options are what to check first.

FAQ

Contextick is one of these stores. You connect it to the assistant you already use, say "remember this", and later, when it's relevant, that assistant brings it back — here's how to connect it.