You set up your AI assistant carefully, gave it your tone, explained your audience. You also gave it your offer, your name, the way you like to communicate. It has everything it needs.
Then the next day, or even the next session, it greeted you as a generic assistant with no memory of any of it.
That is not a glitch. That is how AI models, and by default AI assistants work.
Understanding why AI assistants forget your instructions is the first step to actually fixing the problem. And the fix is simpler than you would think.
Why AI Assistants Lose Your Instructions
AI assistants do not have memory the way we do. When you start a new conversation the model starts fresh. It has no record of what you told it last week, last session, or even an hour ago. Whatever context you built up is gone.
Even within a single session, there is a limit to how much information the model can actively hold onto. This is called the context window, which is the working memory of the conversation.
Think of it like a desk. Once the desk gets full something has to fall off the edge to make room. When the desk overflows, the AI drops older information, and often that older information contains the instructions you first gave it.
This is why it seems to “forget’. You set the context early in the conversation, but the conversation got long and by the end the AI Assistant is working without that context and the output reflects it
This Happens in Professional AI Builds Too
This problem is not unique to beginners using AI assistants, it happens to everyone at every level.
When building Samaritan, a multi-agent OSINT automation system, one of the earliest and most stubborn problems was that it kept forgetting it’s name and who I was. Even though identity and operating rules were loaded at the start of each session. The platform was supposed to read that file and launch the agent with a defined set of behaviors. Instead, it kept coming back as a generic, cheerful, directionless AI assistant with no trace of what it had been given.
It took me three days to figure out what was happening, and back then (a few months ago) everyone was still trying to figure OpenClaw out
Part of the problem (I later discovered) was mistakes that I made with the build. I had multiple copies of the instruction file scattered across different directories, none of them in the location the platform was actually reading.
On top of that, the instruction file was too large for the model’s effective context window, which caused the platform to silently trim it before the model ever processed it. The trimming happened without any notice or error message, which made troubleshooting it even harder.
- Recommended Length of your Soul.md or Claude.md files is 30 to 80 lines (or about 400 to 500 words)
Every token you put in your
soul.mdis processed on every single turn. If your file grows too large (past 1,200 words), your agent’s reasoning window gets cluttered, which can cause it to start missing core instructions or producing lower-quality outputs – Roberto Capodieci
So even a professional build, running on dedicated hardware, failed for the same basic reason your ChatGPT output sounds generic: the model did not have access to the context it needed, and there was no warning that this had happened so there was no way to know.
What You Can Do About It
Depending on which AI Assistant or Agentic AI platform you’re using, you have a few options, and none of them require technical expertise.
Use the memory or custom instructions feature.
ChatGPT and Claude both have ways to store persistent context. In ChatGPT, this is the custom instructions section and the memory feature.
In Claude, you can set preferences directly. Anything you store there gets loaded into every conversation automatically. This is the first thing to set up if you have not already.
Repeat your context at the start of important sessions.
If you are doing something that matters, do not assume the AI Assistant knows your situation. Paste in a short brief: who you are, who you serve, your tone, and the goal of this specific session. It takes thirty seconds and it changes the output quality significantly.
Keep your instructions short.
Long context blocks can hit the same problem over and over. Every model has a working memory limit. If your instructions fill it before the actual work starts, something gets dropped. Three to five focused sentences outperform two paragraphs of context every time.
Save a master prompt document.
Create a simple document with the core context you use regularly: your business description, your audience, your voice, your most common tasks. When you start a session, paste it in. When you update your business, update the document. Treat it like a template.
Use projects
Both ChatGPT and Claude have projects, separate little work areas with special instructions and documents that you can upload just for that specific project. ChatGPT connects to your Google Drive, Claude lets you upload the documents directly.
Remember to convert your docs to markdown. AI models read markdown easier and quicker than text in any other format. The harder it is to parse your documents, the more tokens you burn, the more time it takes, and the more likely it is to skip important information.
They are better at reading PDFs than they used to be, but for your projects try to use markdown or plain text instructions.
Your AI Assistants are not broken
Watching AI assistants ignore what you told it feels like the tool is not listening. But it is not ignoring you. It is working exactly as designed, and the design is that it does not carry memory forward unless you build that memory in.
You do not need to rebuild anything. You just need to stop assuming the AI model knows what you know and start giving it what it needs at the start of every session.
That one change will get you closer to consistent output than any other adjustment you can make.
AI Consulting and Support Specialist
Sec+ CySA+
If you and your team are struggling to integrate AI strategies into your business, I can help. Let’s do a free 30 min chat via Google Meet and see if we can start turning your AI problems into solutions.
