Where each kind of information goes

An agent can be given six kinds of things. They look alike, and people mix them up. This page says which one to use for what.

The comparison

You want the agent to…	Where it goes	How the agent gets it	How long it is kept
Always follow some rules, a tone, a role	Its instructions, in the window of the AI Agent	It reads them at every request	As long as the agent
Have instructions written and tried before	The Prompts library, in Prompt Studio	It does not read it. You copy the result into its instructions	As long as the Prompt
Look up or change facts kept in rows: prices, stock, orders, customers	A Data Table	A Data Table tool	Until the rows are changed
Know your texts: policies, guides, manuals, a website, a feed, the rows of a database seen as texts	A Knowledge Base	The Knowledge Base step, on the diamond Knowledge Bases. The agent searches by meaning	Kept, and kept up to date from its source
Remember the earlier messages of a conversation	An Agent Session	The diamond Agent Session	In-Memory: until the server restarts. File: kept
Keep texts while it works and find them again by meaning: notes, what it learned	A Vector Store	The AI Agent Vector Store Tool, or a Vector Store step in a workflow	In-Memory: the time of one execution. Qdrant, Pinecone, pgvector: kept

An example, with the agent Reminder Writer

The agent Reminder Writer writes a payment reminder to a customer.

What the agent needs	Where it goes
”Write a short, polite e-mail. Say the invoice, the amount and the date it was due.”	Its instructions
The open invoices of the company: customer, number, amount, due date	A Data Table `invoices`
The payment terms and the reminder policy of the company, written as texts	A Knowledge Base `Payment terms`
What was said earlier with the person who is asking, in this conversation	An Agent Session
A note the agent wants to keep: “Northgate Bakery prefers to be called before 10”	A Vector Store, with the AI Agent Vector Store Tool

Which one?

Knowledge Base or Data Table

A Data Table holds rows: the same fields for each line. The agent looks up a value, or changes it. “What does invoice INV-2041 amount to?”
A Knowledge Base holds texts. The agent finds the passage whose meaning fits. “What do our terms say about a late payment?”

If it fits in columns, use a table. If it is prose, use a base. A base can also read the rows of a database, but it turns them into texts: the agent finds them by meaning, and cannot change them.

Knowledge Base or Vector Store

A Knowledge Base is what the agent must know. You prepare it, it reads a source, and it stays up to date.
A Vector Store is what the agent collects while it works: notes, answers found earlier. Something puts texts in it, a step or the agent itself, and something finds them later.

A Vector Store is never the way to give an agent its knowledge. Give it a Knowledge Base.

Agent Session or Vector Store

An Agent Session keeps the conversation, message by message, so that the agent remembers what was said just before.
A Vector Store keeps texts found by meaning, whenever they were put there, whatever the conversation.

Instructions or Prompts library

The Prompts library is a workbench. You write a prompt, try it, improve it. No step and no agent reads it: when it is good, you copy it where it is used. See What Prompt Studio is for. Not to be confused with the step LLM Prompt, which is a step of a workflow, with a prompt written in it.

What the Assistant does

The Assistant of the Studio follows the same rules when it builds a solution for you. It puts rows in a Data Table, texts in a Knowledge Base, and the conversation in a session. It never uses a Vector Store for what an agent must know.

The Assistant cannot create a Knowledge Base. When your texts are in no base yet, it tells you which base to create and what it should read, and it plugs the base in once it exists. See What a Knowledge Base is for.

Go further

Next step

Words we use.