Memory (AI Agents)
Memory refers to the ability of an AI program to retain information from previous conversations and reuse it later. Without this function, the program starts from scratch with every conversation.
A chat program like ChatGPT normally forgets everything as soon as a conversation ends. Next time, it no longer knows your name or what you were working on. Memory is the technique that changes exactly that. Important information is stored outside the conversation and retrieved again when needed. The English term has become established usage; what is meant is a memory. It is especially important for programs that independently carry out multi-step tasks. Such programs are called AI agents.
Why forgetting makes agents useless
A single chat works fine even without memory. You write everything necessary into the question, and the answer fits. But once a task runs over days, this becomes impossible. An agent planning a travel itinerary needs to know what it has already booked the day before.
There is a hard technical limit. A language model can only take in a limited amount of text at once. This intake window is called the context window. It has become large by now, but it is not infinite. Anyone who writes a hundred conversations into it in full will blow past it.
On top of that comes the cost. Every word written into the context window costs computing time and thus money. A good memory system therefore doesn’t deliver everything, but only the three or four relevant notes. This is precisely the economic core of the matter: an agent with memory is not only smarter, but also cheaper to run.
Notepad, archive, and search function
Technically, memory almost always consists of two parts. The first part decides what gets stored at all. After a conversation, the model writes itself short notes, such as: user codes in Python. These notes end up in a database outside the model. The model itself does not change in the process.
The second part is search. Upon a new question, the system searches its archive for matching notes. This usually happens via semantic similarity rather than exact words. To do this, every text is converted into a long series of numbers that describes its content. Texts with similar meaning produce similar number series and can thus be found. The hits are then placed in front of the actual question.
Experts often distinguish three types here. Short-term memory is the ongoing conversation. Long-term memory consists of permanent facts about the user. Episodic memory stores concrete past events, i.e. what the agent did and when. A common misconception, by the way, is that memory means the model keeps learning. That is not true. The model remains unchanged; it merely reads its own notes afterward.
Memory in products and in the headlines
The function is most visible in chat programs. ChatGPT, Gemini, and Claude now offer to remember things about the user. In the settings, you can view and delete the stored notes. Coding tools also use memory to remember a team’s project structures and preferences.
In business news, the term comes up for two reasons. First, memory is considered a prerequisite for agents to take on real work, for example in customer service over the course of weeks. Second, it is a retention tool. A provider that has known you for a year is harder to switch away from than one that doesn’t know you.
Data protection is the other side of the coin. A memory about users is a collection of personal data and falls under European data protection rules. Incorrect notes are another problem. If a system remembers a piece of misinformation, it can repeat it for months. That is why a delete button belongs in every serious memory system.