Customers hate repeating themselves. That is the case for giving an AI agent a memory. But everything an agent remembers is data your business is holding about a person, so the real question is not technical. It is a policy decision: how long should this context matter, and who should be able to see it?
Three answers, for three kinds of work
| How long it remembers | Good for | Watch out for |
|---|---|---|
| Nothing after the task | Checking a document, answering a one-off question | Customers repeating themselves if they come back |
| A day or so | Support chats and assistants that pick up again | Long gaps between replies |
| Until you delete it | Long customer relationships, regulated records | Privacy rules, deletion requests |
Nothing kept
If the agent finishes its job in one go, such as reading a claim or answering a single question, it does not need to remember anything. This is the safest choice: there is nothing to leak later because nothing is kept.
A day or so
Conversations pause and pick up again. A memory that lasts a day, and resets each time the customer writes, keeps active conversations going while old ones fade away on their own.
Until you delete it
Some histories must outlive the conversation: a customer's past requests, or an interaction a regulator may ask about. Keeping a full history is useful, but it means you need a way to delete a person's history when they ask.
Who the memory belongs to matters as much as how long
Memory has to be tied to the right person and the right conversation. Tie it too broadly, and one customer's details could surface in another's conversation. Tie it too narrowly, and customers repeat themselves across channels. Decide whether history follows the customer, the conversation, or a single task.
More memory is not always better
Everything an agent remembers is read again before every reply. That adds cost and can pull the agent off topic. Keep the recent conversation, and let the agent look up lasting facts, like a customer's plan or open orders, from your own systems when it needs them.
Remembering conversations is different from knowing your documents
An agent remembering what a customer said is not the same as an agent knowing your policies or product catalog. Those belong in the agent's rules and in your systems, where they are kept current and apply to every customer equally.
How Aether handles it
Each Aether agent's memory is chosen for its job: nothing kept after the task, a conversation kept for a day by default, or a customer's history kept until you delete it. Your policies and product rules live separately, in Calibration, and what is recorded about each task can be limited to what your data policies allow. See Security and data.