Session: half an hour. You need: a conversation still open on work that's been going on for a few days, and the tool you normally use.
Today you ask your tool four questions, one for each of the places that what it "knows" can come from. The point is to stop asking it to remember things it can't remember, and to know where to look when it gets something wrong.
1. What it has always known
Open a new chat and ask something general about your line of work, without giving any context: "What is a tender specification and what does it usually contain?". It answers straight away and without sources, because that knowledge sits in the weights: it learned it during training. Then ask: "Up to what date do the things you know without searching go?". The date it gives you, even if approximate, is the boundary beyond which, without a search, it knows nothing.
2. What it rereads
Go back to the long conversation and write: "List the decisions we took at the start of this conversation, in the words I used." Then open a new chat, outside any project, and ask the same question. In the first it finds them because it rereads the whole chat with every message. In the second it doesn't have them in front of it. If in the first it gives you back a thinner version, or skips one, you've seen the chat filling up from stage 9. If in the second it finds them anyway, it took them from the product's memory or from a search of past conversations: that's the next step.
3. What the product passes to it
In the new chat, write: "What do you know about me? List everything, one line per item, and for each item tell me whether I told you or you inferred it." Then open the memory settings (where they are, you'll find below, in the section for your environment) and compare the two lists. What you see there is the profile the system puts in front of the conversation before you write. If you use a project, ask the same question inside the project: it also answers with the documents and instructions you've put there, and outside the project it doesn't see that material.
4. What it goes and looks up
Ask something it can't know from its weights: "What has changed in the last week on [a topic in your work]? For each point give the source with the link and the date." An answer that comes from a search brings links to specific pages and recent dates. Open at least two and check that they say what it reports. An answer without links about facts from this week should be treated as a guess.
What you just saw
When the assistant "knows" something, the information can come from five places: from the model's weights, from the conversation in front of it, from a profile the product passes to it before your message, from a project's materials, from a search done at that moment. The table lines them up.
| Memory | Where it sits | How you recognise it | What you can do with it |
|---|---|---|---|
| Structural | In the model's weights | Answers straight away, without sources, and stops at a date | For recent things, ask for a search |
| Conversational | In the open chat | Disappears when you open a new chat | One chat per topic, the handover note (stage 9) |
| Persistent | In a profile kept by the product | Knows who you are even in a new chat | Read it, correct it, clean it up (stage 11) |
| Project | In the project's materials | Knows certain things only inside the project | Choose what to put in it (stage 12) |
| External | In the tools it consults | Brings links and dates | Open the sources (stage 15) |
When an answer is wrong, the first useful question is which of these five it comes from: out-of-date information in the weights, an instruction lost in a chat that's too long, a profile item that no longer holds, a source read badly. Each has a different remedy.
Chapter 5, "Getting what you need", explains the five memories in the section "The memory of LLMs", with the example of a single request that uses all of them at once. The section that follows, "How to make the AI remember you", is the starting point for stage 11.
In the book. Chapter 5, section 5.3 "The memory of LLMs", page 98. The book on Amazon.
Memory is turned on and read in Settings > Memory, with the option "Generate memory from chats". It's available on the Free, Pro and Max plans, on the web, in the desktop app and on mobile; on the Team and Enterprise plans, whoever administers the account decides. What Claude remembers is divided by topic, and each one can be opened, corrected or deleted. Each project has its own memory, separate from the others: that's why at step 3 you get different answers inside and outside the project. Searching past conversations is only available on paid plans, and it doesn't retrieve incognito chats.
Memory is in Settings > Personalization, in the "ChatGPT memory" section: a switch, "Enable memory", and a "Memory summary" with the Manage button, which shows what ChatGPT has learned about you; from there you can also manage memories saved in the past. A saved memory is deleted from the settings or by asking it to forget it, and a temporary chat neither creates nor updates memories. Projects are available on the free plan too, with five files per project, and they can be set up with project-only memory: the project's chats see the project's other chats and nothing else. According to OpenAI's help pages, memory features vary by plan, country, platform and organisation settings: what your account has, your settings will tell you.
In Copilot Chat, memory is found from the three dots at the top right, Chat settings > Personalization, under "Saved memories". It brings together three things: saved memories, details inferred from chat history and custom instructions. It's on by default, even without a Microsoft 365 Copilot licence, but the administrator can turn it off for the whole organisation: if you find the switches off and you didn't turn them off, the decision belongs to the IT department. To make it keep something, just ask it in the chat; to remove it, delete the item from saved memories. The memories sit in your work mailbox, in a hidden folder. Microsoft describes the feature as a preview and subject to change. Instead of a project there's the Notebook, which bases its answers, selectively, on the sources you put in it.
All four sources are there. What it has always known: the weights of the model you choose from the model picker. What it rereads: the conversation and the uploaded files. What the product passes to it: the "Personal context", that is, the memory of previous chats and your instructions (stage 11), and the connected Google apps (stage 16). What it goes and looks up: Google Search, with the links at the bottom of the answer. When an answer has used memory, a label appears at the bottom among the sources.