Phase 0 · Foundations
Context and memory: what the model can see
By the end, you'll be able to…
- Explain what a context window is and why long chats start to drift
- Tell the difference between a single chat's memory and a tool's persistent memory feature
- Decide when to start a fresh chat instead of pushing on in a muddled one
Why it matters
AI tools feel like they 'remember' you, and sometimes they do, but not the way you'd think, and not always. Knowing exactly what the model can and can't see in front of it explains why answers get worse in long conversations, why it sometimes forgets what you told it two messages ago, and when a clean restart will fix everything.
What "the model saw" really means
In the last lesson you learned a model predicts the next token from the text in front of it. The natural question is: what text, exactly, is in front of it?
The answer is: everything that fits inside its context window, the amount of text the model can hold in view at once while it writes a reply. Think of it as the model's working desk. Your latest message, the earlier back-and-forth in this chat, any files you've attached, and the tool's own hidden instructions all sit on that desk. When you send a message, the model reads the whole desk and predicts what comes next.
The context window is measured in tokens and it is large but finite. Modern tools hold a great deal, often the equivalent of a long document or a whole meeting's worth of chat. But it's not infinite, and that single fact explains most of the surprising behaviour you'll meet.
Why the model "forgets" things you said
If everything on the desk is visible, why does a tool sometimes forget an instruction from earlier?
Two reasons. First, the desk fills up. In a very long conversation, the oldest messages can fall off the far edge of the context window to make room for new ones. The model isn't ignoring you; it can no longer see what you said an hour ago. Second, a full desk gets cluttered. Even when everything technically still fits, a model asked to weigh a huge, sprawling conversation reasons less crisply than one given a short, focused one. Important instructions get diluted by everything else competing for attention.
That's why long chats degrade. A conversation that started sharp turns woolly: it repeats itself, loses the thread, contradicts an earlier answer, or forgets the format you asked for. It's not getting tired; its desk is just overloaded.
When to start fresh
The fix is almost always the same and almost always underused: start a new chat. A fresh conversation is an empty desk. If you're switching topic, or a chat has grown long and started to drift, open a new one and, if needed, paste in just the handful of things that actually matter. You'll often get a noticeably better answer in the first reply of a clean chat than after ten messages of nursing a muddled one.
A good rule of thumb: one task, one chat. Keep the conversation about redrafting a policy separate from the one about analysing sales figures. It keeps each desk clean and makes your history easier to find later.
Chat memory versus persistent memory
Here's the distinction that trips everyone up.
Within a single chat, the model "remembers" earlier messages only because they're still on the desk, sitting in the context window. Close that chat and open a new one, and none of it carries over. That's chat memory, and it lives and dies with the conversation.
Separately, most 2026 tools now offer a persistent memory feature: an opt-in setting where the tool saves specific facts about you, your job, your writing style, projects you keep mentioning, and quietly slips them onto the desk in future chats, even brand-new ones. So if it greets a fresh conversation already knowing you work in operations at a mid-size firm, that's persistent memory at work, not the model magically recalling an old chat.
The two behave very differently, and it's worth keeping them straight:
- Chat memory is automatic, per-conversation, and gone when the chat ends.
- Persistent memory is a deliberate feature you switch on, it spans conversations, and you can usually view and edit exactly what it has stored.
The shared 2026 concepts you'll meet everywhere
Every major tool now wraps context and memory in three features. We'll cover each tool's version later; for now, just learn the ideas so the settings menus make sense.
- Custom instructions. A short, always-on note you write once: who you are, how you like replies (concise, British spelling, no jargon). The tool puts it on the desk at the start of every chat, so you stop repeating yourself.
- Persistent memory. As above: saved facts about you, carried between chats, viewable and editable.
- Projects or workspaces. A named container that keeps a set of chats, files and instructions together, so everything about, say, the office-move decision shares one clean context instead of scattering across your history.
A worked example: the long chat that lost the plot
Picture a long single chat at Fernway where you've been drafting a customer-feedback plan. Twenty messages in, the tool starts ignoring the "keep it to one page" instruction you gave at the top; it's fallen off the edge of the desk.
Rather than argue with it, you start fresh and hand the new chat exactly what it needs:
I'm drafting a customer-feedback improvement plan. Keep the whole thing to one page. Here are the three problems we've identified: [paste the three points]. Draft the plan now, in that one-page limit.
Why this works: A new chat is an empty desk. Restating the goal, the constraint and the material in one message gives the model a short, sharp context instead of a cluttered twenty-message one.
The first reply of the clean chat respects the limit again, because the "one page" rule is now sitting right at the front of the desk instead of buried under twenty messages.
To stop repeating that constraint in every chat, you'd set it once as a custom instruction:
Custom instruction: I work in operations at a mid-size UK company. Default to concise replies in British English. When I ask for a document, keep it to one page unless I say otherwise.
Why this works: A custom instruction rides onto the desk at the start of every conversation, so a preference you'd otherwise retype each time becomes automatic.
And when you want the model to lean on what it saw earlier, ask for it plainly:
Using only the meeting notes I pasted above, list the actions that were left unassigned. Don't invent any that aren't in the text.
Why this works: Explicitly referencing 'the notes above' tells the model which part of its context to work from, instead of guessing which thread you mean.
Try it now
Common mistakes
- Piling every topic into one endless chat. The desk overflows, old instructions drop off, and answers drift. One task, one chat.
- Blaming the model for "forgetting". It didn't lose your instruction out of carelessness; it literally fell out of the context window. Restating it or starting fresh fixes it.
- Confusing the two kinds of memory. Expecting a brand-new chat to recall an old conversation (it won't, unless persistent memory saved a fact), or being surprised a fresh chat "knows" you (persistent memory did that).
- Over-trusting a long chat's later answers. As a conversation grows, the model is more likely to contradict something it told you earlier or quietly drop a constraint. Don't assume message twenty is as reliable as message two; re-read the key output rather than trusting that it "remembered".
Keeping current
Context-window sizes, memory features and how projects work change often and differ by tool and tier. When it matters, check the tool's own documentation on memory and context, for example OpenAI's memory help article, rather than assuming last year's limits still hold. Correct as of 13 July 2026.