The basics
What is AI memory?
AI memory is a store of short facts kept outside a language model. Before each answer, the facts that matter for the question are found and added to the prompt. That is all, and it changes what an agent can be.
The problem it solves
A language model has no state. Each request starts from zero, and the only things it knows about you are the words in the prompt in front of it. Chat products hide this by sending the whole conversation again on every turn. That works until the conversation ends, or grows longer than the model can read.
Memory is the fix. The agent writes down what is worth keeping, stores it under your name, and reads the relevant lines back before it answers, today or in three months.
Why a basket
In the hills of Nepal a doko is a woven bamboo basket carried on the back. It holds a lot, but nobody fills it with everything they own; they pack for the journey. An agent's memory works the same way. The store can be large. What goes into the prompt for one question must be small.
The types of agent memory
Short-term memory
The current conversation, held in the context window. It is exact, and it is gone when the session ends or the window overflows.
Long-term memory
Facts kept in a store outside the model and brought back when needed. It survives across sessions, devices and model upgrades.
Semantic memory
Things that are true: "is vegetarian", "the project uses PostgreSQL".
Episodic memory
Things that happened: "on 3 June the deploy failed because of a missing migration".
Procedural memory
How to do things: "run the tests before every commit", "always answer in bullet points".
What AI memory is not
Not training
Nothing is learned into the model's weights. Delete a memory and it is gone; you cannot do that with training data.
Not a bigger context window
A window is one request. Memory is what is still there for the next one.
Not a chat log
A log is everything that was said. Memory is the small part of it worth carrying.
Not document search
RAG looks things up in documents written beforehand. Memory looks things up in what the agent itself has learned.
AI memory: common questions
What is AI memory in simple terms?
It is a notebook an AI agent keeps about you or about a task. Before answering, it reads the lines that matter for your question. After the conversation, it adds anything new worth keeping.
Do large language models have memory?
Not on their own. A model receives text and returns text, and keeps nothing between requests. Any remembering is done by software around the model.
Where are AI memories stored?
In an ordinary database outside the model, most often one that supports search by meaning using embeddings. Some systems add a graph of entities or plain structured records.
Can an AI memory be deleted?
Yes, and it should be easy. Because memories are rows in a store and not model weights, a single fact or a whole user can be removed.
Does memory make answers more accurate?
It makes them better informed about the person or task. It can also carry forward a wrong or outdated fact, which is why updating and forgetting matter as much as storing.
Pack the doko. Ask it anything.
Nine memories, one question, no account. See which facts an agent would carry into its next answer.