AI memory vs the context window, explained

The context window: what the model sees this turn

Every time you send a message, the model gets a big block of text to work with. That block is called the context window. It contains the system prompt (invisible instructions from the provider), the chat history so far (as much as fits), any tool output, and your latest message. The model reads all of it, then generates a reply.

The context window is finite. Different models have different sizes — GPT-5 and Claude 4.5 both handle very large windows in principle, but the practical size in the chat UI is smaller because the provider stuffs a lot of scaffolding in first. When the window fills, older messages get pushed out. This is why "message 400 in a long chat feels like the AI forgot message 12" — because it did. Message 12 isn't in the window anymore.

Memory: a separate layer that survives between windows

Memory is not the context window. Memory is a separate storage layer — usually a small set of notes the model has decided to write about you — that persists across chats. When you start a new chat, that new chat has an empty context window, but the memory system can inject the most relevant notes back into that window so the model has a starting point.

Two systems, two failure modes:

• The memory system didn't write the fact (nothing to retrieve later).
• The memory system wrote it, but didn't inject it into this turn's context window (so the model can't see it, even though it's stored).

Almost every "AI forgot me" complaint is one of these two, separately or together.

Why bigger context windows aren't a memory upgrade

When providers announce "1 million token context window" it sounds like memory got 100× better. It didn't. It got 100× better at holding a single long thing in mind — a giant document, a long debugging session — but a new chat with a million-token window still starts empty. Nothing persists to the next chat unless a separate memory layer stores it.

There's also a subtle wrinkle: even inside a huge context window, the model doesn't weigh everything equally. The recent turns and the very first turns tend to have the most influence; the middle often "gets lost." So even with a giant window, quality of recall across a very long chat is worse than the raw number suggests.

What each is good at

The practical takeaway

If the frustration is "in this chat the AI is forgetting things from earlier" → context window problem. Fix: start a new chat, summarize the essentials, paste them in fresh. Or use a Project / Space that keeps a system-instruction layer.

If the frustration is "in a new chat the AI acts like it doesn't know me" → memory problem. Fix: check memory settings, check the saved-memory panel, and — for the durable version — keep a portable persona doc outside any one provider so any new chat can start with the essential context pre-loaded.

Diagnosing which problem you have is the fastest path to fixing it. See ChatGPT memory not working for the memory-side playbook, and how AI forgets you for the deeper mechanics.