Explore

How an AI assistant works, acting for you across your accounts

Point an agent at one person's inbox and calendar. Each step adds a piece, from memory to tokens, and shows what can go wrong with it. About 10 minutes.

Step 01 of 06

What it remembers about you

Here's the agent from “How an AI agent works, one piece at a time”: the app (the chat window plus the code behind it, which sends text to the model, runs tools, checks permissions and keeps memory), the model (it predicts text from what the app sends it and never runs anything) and the two in a loop with tools, the agent. Now the user is one person: Dana, a sales lead at a 40-person software company. Her assistant is that app and model, run by its maker, a lab such as OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini) or Microsoft (Microsoft 365 Copilot). Google, her account provider, holds her mail and calendar, and other people email her. First, memory. A saved memory is a short note about Dana that the assistant sends the model at the start of every chat. She can ask it to remember something, and it can save a note on its own. It lives on the maker's servers, apart from her chat history.

Your guess

In March, Dana mentioned in passing that Sam's team is her biggest account, and the assistant saved it. Later she deletes that chat. Does the assistant still know?

Make a guess to continue.

The answer is A. OpenAI's help center says deleting a chat doesn't remove a memory saved from it. The note lives apart from the chat, and goes out with every new one.

What can go wrong now

A memory stays true for the assistant after it stops being true for Dana, and she may never have seen it being written. The note was the model's choice from text it read, the same channel prompt injection uses: in one study, poisoning less than 0.1% of a memory store made attacks succeed more than 80% of the time.

How teams handle it

Decide up front whose memory it is, how long an entry lasts and how it gets erased. Show every saved memory where the person can edit it, and treat a memory write as a write action. More in Building applied AI products: context, connectors and memory.

Next:Memory knows about Dana. To do anything with her inbox, the assistant needs her permission: Connecting her accounts.

Questions people ask

Does the model itself remember me?

No. The app keeps the notes and sends them with every chat, as in “How an AI agent works, one piece at a time”. The model sees them the way it sees anything else the app sends.

Is memory the same as training on my chats?

No. A saved memory is a note the maker keeps and sends to the model with each chat, and the model itself doesn't change. Whether chats are used for training is a separate question, answered in each maker's settings and terms.

Can Dana see and control what it remembers?

In ChatGPT she can see and delete saved memories in her settings, turn on reference chat history so it also draws on past chats, or use a temporary chat that leaves memory alone. A company building its own assistant on Anthropic's memory tool keeps those notes in its own storage.