Memory

What Memory is

Memory is what an AI tool keeps from one conversation to the next, so it does not have to be told the same context again. Without it, every conversation starts from nothing and you explain who you are and what you are working on. With it, some of that context is already in place when you begin.

The AI Agents and Personalisation session states the aim as a chatbot set up around your own role, so it stops asking for the same context every time. Memory is one of the two tools for that. Custom instructions are the other.

The remembered half of Personalisation

Personalisation on this site has two halves. Custom instructions are the stable half. They are rules you write yourself, and they stay as written until you change them. Memory is the remembered half. It is what the tool keeps as you work, so it changes over time.

The personalisation outline lists memory among the six ways to personalise a tool. It teaches it in the same block as custom instructions, traits and style, personality and tone, occupation and background, and naming and identity. The two halves are taught together because choosing between them is a judgement about how long a fact stays true.

When AI should remember something for later

The courses treat memory as a judgement and not a switch. The personalisation session has a topic on exactly this, when AI should remember something for later, and the definition of the term on this site includes the other side of it, when it should not.

A fact that holds for every request, such as your role or the format you want back, is a standing rule and belongs in custom instructions. A fact that belongs to one piece of work is context. It may be worth the tool keeping for later, or it may be better placed in the prompt for that task. The question is how long the fact will stay useful, and whether it is safe for the tool to keep it.

How it fits with Custom Instructions

The HR course gives a session to custom instructions against memory and when to use which. The editors and authors session lays the sequence out in order. It teaches custom instructions, then traits, then memory, then when to use which, and only after that moves on to advanced persona design and meta prompting. Standing rules come first, memory second, and the harder techniques after both.

The Mistral workshop covers the same ground as stable preferences against changing context, and closes the point with a judgement. The useful skill is deciding which of your preferences are stable enough to set once and which change with the task. Memory is the place for the second kind, and it is the reason a team needs both halves rather than one.

A Prompt that writes them for you

The personalisation session includes a prompt that writes both your custom instructions and your memory for you, so you do not start from a blank box. Asking the tool to draft what it should keep, and then reading the draft, is a more careful way to fill memory than letting it collect whatever comes up.

A worked example

The personalisation session gives a use case that shows the split. In standardised ticket replies, the tone and the shape of the reply should be the same every time, which is what standing rules are for. What changes from one ticket to the next is the customer and the problem, which is context.

The one-to-one programme adds a learning loop to the same idea. A person’s corrections become dated rules that the tool keeps. A correction made once is not made again, and because each rule carries a date, it can be reviewed later. That is memory at its most practical, and the date is the detail worth copying. A remembered rule with no date is hard to question.

Memory in an Agent

Memory has a second meaning once you build an agent, and the distinction that matters most in practice is one the databases session draws. An agent with a fixed knowledge base answers from what it was given. An agent with a live database it can read and write has memory, which a fixed knowledge base does not.

The first is grounded in a snapshot. The second is grounded in something that changes, and it can add to it. That is why the advanced course has a block called agents with memory. It covers databases and building them with AI, and it shows a prompt library agent. The bespoke Excel and databases session builds a project and prompt library agent live, with version history so the library compounds, and teaches building the database first, before you build the agent. It also covers how agents work with databases and where they struggle.

A team choosing between the two is really choosing whether the material should stay as it was or keep changing as people use it. Retrieval-augmented generation covers the fixed side of that choice, and the AI agent page covers what gains or lacks memory depending on how it is built.

Across the Tools

The idea is not tied to one product. The ChatGPT courses list personalisation and settings as part of the syllabus, and memory sits inside it. The Mistral workshop covers memory as part of personalisation in its own tool. The editors and authors session and the advanced generative AI course both include it, and the advanced course has a personalisation session with custom instructions, memory and personalise your chat.

The site does not describe any product’s settings screen, so a team should check what its own tool keeps and where. One thing the catalogue does say is worth noting. The review of two vendors’ AI offers reports one vendor’s own comparison stating that two of another vendor’s products do not share memory or context with each other. What one product remembers is a reason to check, not to assume, before relying on it in another.

Why AI forgets

The personalisation session closes on the limitations of large language models, and lists why AI forgets, loses context, or needs a stronger set-up. Memory is one answer to forgetting. It is not the only one, and it is not always the right one.

The Mistral workshop closes on a decision framework that asks when to use chat, a project, a library, memory, custom instructions, an agent, a connector, Deep Research or Work Mode. The point is to choose the smallest option that does the job. If the missing context is small and belongs to one task, put it in the prompt. If it is a rule you apply everywhere, use custom instructions. If it is something the tool should carry and add to, that is where memory, or an agent with a database, earns its place.

What not to let it remember

Everything you type into an AI tool is data, and where a tool keeps some of it for later, the same rule covers what you tell it to keep. The site’s article on why banning AI at work backfires argues for a rule drawn by kind of data and not by tool. Client information, contracts and HR records should not go in unless a rule says they may.

The personalisation course makes the same point in its closing topic, using public AI safely, with public, anonymised, generic or invented content only. Memory is convenient because it holds context for you. That is also why it is worth deciding, before anyone starts, what a person is allowed to tell a tool that will keep it.

Mix-ups worth avoiding

Four confusions come up often enough to name. The first is treating memory and custom instructions as one thing, when one is written by you and stays fixed and the other is kept by the tool and changes. The second is treating a fixed knowledge base as memory, when a fixed one is a snapshot and only a live database that the agent can read and write gives it memory.

The third is assuming a tool that remembers you is a tool that is right. Memory saves you restating context, and it does not check a fact. The fourth is assuming that what one product remembers will carry over to another.

Where it goes wrong

Memory does not check facts. A tool can remember your role, your tone and last week’s project, and still produce a reply that reads well and is wrong. That is hallucination, and it is a property of how generative AI works, not a fault in what the tool remembers.

The working habit does not change. Ask for the source, open the source and treat the first reply as a draft. The person who sends the result on is responsible for what it says. A tool that remembers you makes this easier to forget, because the reply feels familiar, and that is the moment to read it most carefully.

Where it is taught and where it appears in an Audit

The AI Agents and Personalisation session is where memory and custom instructions are set up together. The HR course teaches when to use one against the other, and the editors and authors session places memory in its sequence. The Mistral Le Chat workshop covers it for its own tool. Six of the courses are the core AI courses, and the rest sit in the wider training catalogue.

An audit meets memory in two places. The first is repetition. Where a team briefs the tool from scratch every morning, the work repeats, and the prompting workshop’s decision framework sorts it by how often it repeats and how much judgement it needs. The second is the choice between a snapshot and something that changes. Whether a workflow should answer from material that stays fixed or from a database that people keep adding to is a design question, and it is settled by what the process handles.