Why your AI forgets everything, and how to give it a lasting memory
You have already explained your business, your clients and your methods to your AI. Ten times. Here is why it doesn’t remember, and what really works so you never have to start again.
The problem: every conversation starts from scratch
You open a new conversation. You ask your assistant to draft a sales proposal. It comes back with a text that is correct, polite and completely generic. It doesn’t know that you sell to local authorities, that you write to your clients informally, or that you never offer a discount in the first year. Yet you told it all that last week.
So you paste it in again. The paragraph that introduces the company, the three rules on tone, the sample proposal that went down well. Five minutes, sometimes ten. And you will do it again tomorrow.
This is not a flaw in your particular assistant. It is how these tools are built. The good news: once you understand the mechanism, you know where to act.
Why an AI has no memory by default
An assistant such as ChatGPT, Claude or Gemini runs on a language model. That model “retains” nothing between two exchanges. With each message, it receives a block of text, reads all of it and produces a reply. That block is called the context.
The context holds everything the model can take into account at that moment: the tool’s general instructions, the history of the current conversation, the files you have attached, and your latest message. Nothing else. Anything that is not in the context does not exist for the model.
Three consequences follow.
- A new conversation is a blank page. The history of the previous one is not reloaded.
- The context has a size limit. It is measured in tokens, which are fragments of words. When a conversation gets very long, the earliest exchanges end up being summarised or dropped.
- Everything in the context is reread with each message. The more text you paste, the more each reply costs and the more likely the model is to miss what matters.
So when people say an AI “remembers”, it always means the same thing: someone, or something, put the right information back into the context at the right moment. The whole question is who takes care of that, and how.
The fixes you are probably already using, and their limits
Pasting the context in by hand
This is where everyone starts, and it works. You keep an “about us” file and paste it in at the start of each conversation.
Its limits show up quickly. The file grows, you paste all of it “just in case”, and the assistant reads three pages to answer a question that needed two lines. You also have several versions lying around, and you no longer know which one is current.
Custom instructions
Most assistants offer a field where you can write who you are and how you want to be answered. That text is added to every conversation.
It is very useful for tone and a few stable rules. But space is tight, and the field is not designed to hold your methods, your prices, a client’s history or the brief for a project.
The assistant’s built-in memory
Some assistants pick things up by themselves over the course of your conversations: your line of work, your preferences, a project in progress.
It is convenient, with three reservations. You don’t choose exactly what is retained. What it keeps are short facts, not working documents. And that memory stays locked inside a single tool: what ChatGPT has retained, Claude doesn’t know, and neither do your colleagues.
“Projects” and attached files
Project spaces let you upload documents that an assistant will consult in every conversation within that project. This is real progress: your documents are there without your having to paste them in again.
The limits are a matter of scope. The project lives in a single assistant, with a single provider. If you switch tools, or a colleague uses a different one, you have to start all over again. And when a document changes, you have to remember to replace the file everywhere it was uploaded.
| Fix | What it does well | Where it stops |
|---|---|---|
| Pasting by hand | Simple, immediate | Slow, multiple versions, everything is reread every time |
| Custom instructions | Tone and stable rules | Too short for working documents |
| Built-in memory | Nothing to do | Little control, short facts, a single tool |
| Projects and attached files | Documents available without pasting them again | A single assistant, manual updates |
| Connected knowledge base | A single source, every assistant, up to date | You have to write things down properly once |
What a lasting memory is
A lasting memory is a place that meets four conditions.
- It sits outside the assistant. Your knowledge doesn’t depend on a provider. You can change AI, or use two, without losing anything.
- The assistant looks things up there by itself. You no longer paste anything. You ask your question, and it consults the source.
- It reads only the relevant passage. Not the whole folder: the paragraph that answers the question. That is what keeps answers precise and usage reasonable.
- It is kept up to date in one place. You correct a method once, and the new version is the one that is read from then on, everywhere.
Technically, this has become simple since the arrival of the MCP protocol, an open standard that lets an assistant connect to an outside source. If the subject interests you, we explain it in MCP explained simply.
Building your memory in five steps
1. Spot what you keep re-explaining
For one week, make a note every time you paste or rewrite the same thing for your AI. The list is always shorter than you expect: who you are, how you write, how you carry out a given task, what has been decided on a given project.
2. Write it down once, properly
One subject per document. A title that says what it contains. Subheadings. Full sentences, with the reasons: “we don’t give a discount in the first year, because the price already covers the set-up”. An AI applies a rule far better when it knows the reason behind it. We set out the method in Building a knowledge base for AI.
3. Organise by project
Group together what belongs together: one project for your general methods, and one project per client or per job. The structure helps you as much as it helps the assistant, and later on it lets you open one project to someone without opening the others.
4. Connect your assistant
Connect your AI to this workspace. With Linkevio, you paste your workspace address into your assistant, authorise the connection, and you are done. The step-by-step instructions are in Connecting an assistant.
5. Run the test
Open a fresh conversation and ask a question whose answer is in your documents, without pasting anything. If the answer follows your method and cites the right document, your memory is working. If not, the document probably has a poor title or is too vague: fix it and try again.
The most common mistakes
- Importing everything at once. You end up with a memory full of drafts and out-of-date versions. The assistant will find the wrong answers too.
- Keeping contradictory documents. If two documents say two different things, the AI will pick one at random. Mark the one that is authoritative and archive the other.
- Writing for yourself. “Check with him about the usual process” means nothing to a reader who doesn’t share your memory. Write as you would for a new colleague.
- Never updating. A wrong memory is worse than an empty one. When a rule changes, correct it the same day.
- Forgetting decisions. People readily write down methods, but rarely choices and the reasons for them. Yet those are what stop the AI from suggesting the idea you have already ruled out.
And if you work as a team
Everything above applies to one person working alone. With several people, the problem changes scale: everyone explains the same client and the same product to their own AI, and no one does it in quite the same way. The answers diverge, and so do the mistakes.
A shared memory settles this in one go: one person writes, and every member’s assistant reads, each with its own access rights. That is the subject of our guide Sharing context across your team’s AI assistants.
How Linkevio goes about it
Linkevio is a knowledge base built for this purpose. You write or import your documents, organise them by project, and connect the assistant you already use. Linkevio has no built-in AI: it serves yours.
In practice, when your assistant needs a piece of information, it searches the projects you have access to, receives short excerpts with their location, then reads only the part of the document it is interested in. It can also write there, if you ask it to: record a decision, update a procedure. Every request is checked and recorded in a log, and you can cut the connection whenever you like.
The free plan covers two people and one project, which is enough to run the test described above. Create your workspace, write your first three documents, and see whether your AI remembers them tomorrow.
Frequently asked questions
My assistant already has a “memory” feature. Isn’t that enough?
For your preferences, often yes. For your working documents, no: that memory retains short facts that you don’t choose precisely, and it stays inside a single tool. A connected knowledge base holds your real documents, which you control and which all your assistants can read.
Does it work with several AIs at once?
Yes. That is the point of a memory that sits outside the assistant: any assistant that accepts connections using the MCP standard, such as Claude or ChatGPT, can connect to it. You can use one for writing and another for coding, and they read the same source.
Will I use more tokens?
Generally fewer. The assistant reads the relevant passage instead of rereading everything you used to paste in as a precaution. The saving depends on the length of your documents and on your habits; we give the orders of magnitude in Reducing your token usage.
Are my documents used to train an AI?
Linkevio contains no AI and trains nothing on your content. What your assistant reads, however, is sent to its provider, under the contract you have with them: check its privacy settings.
How long does it take to get started?
Half an hour to write three useful documents and connect your assistant. The rest builds up as you go: every time you are about to re-explain something, write it down instead.