Context & Memory
Why does the AI remember your whole trip plan for fifty messages, then suddenly forget your name from message one? Why does a new chat know nothing you told the last one? Ten minutes here demystifies AI "memory" completely — and turns its quirks into habits that work for you.
How AI "remembers": the context window
Here's the surprise: the AI doesn't remember your conversation the way you do. Every single time you send a message, the tool quietly hands the model the entire chat so far — every message from both sides, top to bottom — and asks "what comes next?" The model re-reads it all, replies, and retains nothing. Next message, same routine: whole transcript, fresh read.
The space available for that transcript is called the context window — literally, the window of text the model can "see" at once. Everything inside the window influences the answer. Everything outside it doesn't exist.
And windows have edges. The context window is huge these days — often hundreds of pages of text — but a long, meaty conversation can still outgrow it. When that happens, the oldest messages fall outside the window. The result is the classic AI moment: fifty messages into planning your daughter's birthday, the AI cheerfully asks how old she is — you told it in message one, but message one has slipped off the edge. It didn't "forget" the way people forget; the information simply isn't in front of it anymore.
Think of the context window as a desk, not a brain. Everything currently on the desk — your messages, its answers, any files you dropped there — is instantly available. But the desk has a fixed size: pile on enough paper and the oldest sheets get pushed off the far edge. And when you walk to a different desk (a new chat), it's bare. Nothing came with you. The skill of this topic is really desk management: keep the important papers on top.
Separate chats share nothing — and why that's good
By default, each conversation is sealed off from every other. The chat where you poured out your business plan and the chat where you're drafting today's email are strangers. This surprises beginners ("I told you this yesterday!") but it's mostly a feature:
- Clean answers. No leftover topic bleeding into a new one — remember the travel-flavoured cover letter from Your First Conversation? Fresh chat, fresh mind.
- Privacy by compartment. Sensitive discussion in one chat doesn't echo into others.
- A better history. One-topic chats give your sidebar meaningful titles you can actually find again next month.
This is the deeper reason behind the "one chat = one topic" rule: you're not just being tidy, you're curating exactly what's on the desk for each job. A chat that contains only rent-negotiation material produces sharper rent-negotiation answers than one that also contains recipes and homework.
Re-priming: the 60-second superpower
So a new chat knows nothing. The fix is a move called re-priming: start the new chat by pasting a short summary of the situation. Sixty seconds of typing buys you a fresh, focused conversation with all the essentials on the desk:
"Context: I run a small online clothing business, 2 years old, mainly selling on Facebook. Monthly revenue is modest but growing. I'm considering opening a physical stall at a weekend market. In this chat I want help deciding whether it's worth it."
Even slicker: before leaving the old chat, ask the AI to write the summary for you — "Summarise everything important from this conversation in a short paragraph I can paste into a new chat." It compresses fifty messages into five lines, and the new chat starts at full speed.
Summarise this conversation so far in under 150 words, as a briefing I can paste into a new chat: the situation, the key facts and numbers, what we decided, and what's still open. Write it so an assistant with no memory of this chat could pick up exactly where we left off.
"But my app remembers my name!" — persistent memory features
Some tools now add a second layer on top of the context window: a memory feature — a small notebook of saved facts about you ("prefers short answers", "lives in Chittagong", "runs a pharmacy") that the tool quietly slips into future chats. That's why ChatGPT or similar apps sometimes greet you with things you told them weeks ago in another conversation.
Worth knowing about this notebook:
- It's selective, not total recall. It saves occasional highlights, not your transcripts. Don't count on any specific detail being in there.
- You can read and edit it. Look in Settings under something like "Memory" or "Personalization" — every saved fact is listed, individually deletable, and the whole feature can be switched off. If an app knowing things across chats makes you uneasy, this is the switch. (More on the privacy angle in Privacy Basics.)
- Its cousin is better for work. If you want the AI to always know certain things — your job, your preferred tone, "answer in Bangla" — the reliable tool is custom instructions: standing orders you write once and every new chat obeys. That's a Chapter 5 topic: Custom Instructions.
Uploaded files sit on the same desk
One more thing occupies the context window: files. When you upload a PDF, a spreadsheet, or a photo, its contents land on the same desk as your messages — which is exactly why the AI can answer questions about page 30 of your contract. Two practical consequences:
- A huge document takes up a lot of desk. Upload a 200-page report and a marathon conversation on top of it may push early parts — of the chat or the document — out of view. For big-document work, keep the chat focused on that document.
- Files don't travel between chats either. New chat, re-upload. (Some tools offer "projects" or workspaces that keep files attached across chats — handy once you're past the basics.)
Four habits that make memory a non-problem
- Front-load the key facts. Put names, numbers, constraints, and the goal in your first message — the stuff from Prompting 101's Context block. Early foundation, everything built on it.
- Restate after a long detour. Twenty messages exploring a side question? Pull the thread back explicitly: "OK, back to the stall decision — remember: budget 30,000, weekends only." A one-line recap re-anchors everything.
- Keep a notes file of context blurbs. A plain note on your phone or computer with your 3–4 favourite intros: one describing your work, one for your business, one for a project. Starting any new chat becomes paste-and-go. This tiny habit is the single best productivity trick in this chapter.
- Watch for forgetting, don't argue with it. If the AI misremembers a detail from long ago, it probably slid off the desk. Don't scold — just restate the fact and carry on. (If it invents a detail you never said, that's a different beast: hallucination.)
Help me write a reusable "context blurb" — a short paragraph I can paste at the start of future chats so you instantly understand my situation. Interview me: ask up to 5 quick questions about my work/life, then write the blurb in under 100 words, in my voice. I'll save it in my notes app.
Treat every chat like briefing a brilliant assistant who has severe amnesia but takes perfect notes while you talk. Brief well at the start, re-brief after detours, and bring your written blurb to every new meeting — and the amnesia stops mattering entirely.
Key takeaways
- AI "memory" inside a chat is the context window: it re-reads the whole visible conversation every time you send a message.
- Long chats overflow the window — the earliest details fall off the edge and get "forgotten". Restate what matters.
- Separate chats share nothing by default. That's a feature: clean answers, compartmentalised privacy, findable history.
- Re-prime new chats with a pasted summary — and let the AI write the summary before you leave the old chat.
- Persistent memory features and custom instructions add cross-chat knowledge; both are viewable, editable, and optional.
- Uploaded files share the same desk — big files, focused chats.