What the AI knows and forgets

01What this module is about

"It forgot what I told it." "It was great at the start, then it went downhill." These are two of the most common complaints about AI, and they have the same cause.

The AI doesn't remember things the way you do. It works from what's in front of it right now, and that space fills up. Once you understand this, you can stop the problem before it starts.

By the end of this module you'll know when to start a fresh chat, how to carry your work across with a short handover note, and how to check what your tool remembers about you.

02Two kinds of knowing

Context: what's on the desk. Everything the AI can see in this chat right now. That's your project files, the conversation so far and any files you've added. This is called the context window. When you start a new chat, the desk is cleared, except for your project files.

Memory: notes in a drawer. Some tools keep short notes about you between chats, like your job or your preferences. Each tool does this differently, and you can't always see what it has kept.

What it never knows: anything you haven't given it. If it isn't on the desk or in the drawer, the AI will guess.

03Why long chats go downhill

The longer a chat gets, the worse the AI gets at keeping track of everything in it. Anthropic, the company behind Claude, calls this context rot. Independent tests on many different AI models found the same thing.

When a chat gets very long, some tools also squeeze the early part into a short summary to make room. The summary keeps the gist, but details get lost. That's often where your rules go.

Signs a chat has gone on too long:

• It repeats a mistake you've already corrected.
• It forgets a rule it was following earlier.
• It contradicts something it said before.
• Its answers get vaguer, or start going round in circles.

The fix is simple. Start a new chat, and bring a short handover note with you.

04Text to copy

At the end of a long chat, paste this:

We're going to carry on in a new chat. Write a handover note for it, under 200 words, with these parts:
What we're working on, and why.
What's done.
What's decided, and why.
What's next.
Anything that went wrong, so the next chat avoids it.
Only use facts from this chat. Don't invent anything.

Then start a new chat inside your project, and paste this with the note underneath:

Here's a handover note from my last chat. Read it, then tell me in two sentences where we are. Then wait for me.

[paste the handover note]

Here's what a handover note looks like for the walking club:

Working on: the March newsletter for Hillside Walkers.
Done: the walks section and the AGM notice.
Decided: no photos this month, because the printer is black and white.
Next: a short welcome for the three new members, then a final check against the Reference file.
Avoid: the last draft gave the River Path walk the wrong meeting time. Always check times against the Reference file.

05Memory in your tool

Memory can help, but it can also carry old or wrong details into a new chat. For a project, your project files are better. You wrote them, you can read them, and you control them.

Here's where to find memory in each tool. Find yours by its name in bold, and skip the others. Menu names change, so look for something close.

• Claude. Memory is on by default for personal plans, including Free. To see and change it, go to Settings, then Your Claude, then Manage. Each project has its own memory. To keep a project fully separate from your other chats, turn off Use account memory on the project page. Incognito chats are never saved to memory.
• ChatGPT. Click your name at the bottom of the sidebar, then Personalization. Under ChatGPT memory, the Enable memory switch turns it on or off. Next to Memory summary, click Manage to see what it has learned about you. Some accounts still have the older screen, with two switches called Reference saved memories and Reference chat history. A Temporary chat doesn't use memory. In a project, open Project settings and change Memory from Default to project-only, to keep the project separate from your other chats.
• Grok. Click your name at the bottom of the sidebar, then Settings, then Data Controls. The switch Personalize Grok with your conversation history lets Grok remember past chats. Next to Memory from your chats, click View to see what it has kept. Chats started with the Private button are never stored. There's also a switch that lets Grok use your X account to personalise its answers.
• Gemini. Go to Settings, then Personal Intelligence, then Memory. Older screens call it "Personal context". It needs a personal Google account, and you must be 18 or over. It isn't available in every country.
• Claude Code. CLAUDE.md, and the files it points to, load at the start of every session. Claude Code also keeps notes of its own. Type /clear to start fresh. When a session fills up, Claude Code summarises it automatically, and early details can be lost. So put any rule that must last into your project files.

07Common problems

• The new chat doesn't know what you were doing. The handover note is missing something. Ask the old chat: "What's missing from this note that the next chat will need?"
• It brings up something you didn't mention. That's memory. Check what your tool remembers, and delete or switch off what you don't want.
• It still forgets a rule. If you need it every time, it belongs in your Instructions file, not in a chat.
• The handover note is too long. Ask for under 150 words, and only what the next chat needs to carry on.

08Keeping it current

• One chat per task. When a task is done, start a new chat for the next one.
• End each work session with a handover note. Update HANDOVER.md, and start next time from it.
• Once a month, check what your tool remembers. Delete anything wrong.
• After you change a project file, start a new chat. An old chat still has the old version on its desk.