What is memory in AI assistants and agents?
Memory lets an AI assistant keep information from past conversations and recall it later. The software around the model holds that memory.
Why does a model not remember on its own?
A language model only sees the text it is given in the current request. Yesterday's conversation does not exist for the model unless it is supplied again. Chat apps feel like they remember because they put earlier messages back into the context window every time. When the window is full, the oldest messages drop out.
Short and long term memory
Short term memory is the part of the current conversation that sits in the window. Long term memory is information carried between conversations: the user's name, preferences, past orders, earlier decisions.
Long term memory usually works like this: the system extracts important facts from a conversation and writes them to a database. In a new conversation it searches for records related to the question (often with a vector database) and gives what it finds to the model. Assistants such as ChatGPT, Claude and Gemini offer this kind of memory to users, who can see and delete what is remembered.
Memory in agents
AI agents work over many steps, so they need memory even more. They have to track which steps are done, what each tool returned and what the user approved. On long tasks an agent writes intermediate results to a notes file or database and reads them back when needed.
Risks
A wrongly stored fact carries over into every later conversation. Storing personal data falls under privacy law: what is kept, for how long and for what purpose must be clear, and users must be able to have it deleted. A well designed memory does not save everything. It keeps what is useful and updates what goes stale.