Knowledge and memory
Attach knowledge to an agent so it can search your documents, and understand what agent memory is and how to manage it.
An agent can read your documents on every run and remember facts between runs. These are two different things: knowledge is material you provide, memory is what the agent learns.
Attach knowledge
In the Setup tab, find the Knowledge section.
Choose Add knowledge
A window opens where you can attach knowledge bases, folders or single files.
Pick the content
Choose from what already exists in your workspace, add files from a connected source such as Google Drive or OneDrive, or Upload a file. To upload, you need a knowledge base to put the file in.
Attach
Choose Attach next to each item. It is saved straight away. Use Detach from agent to remove something later.
Good to know:
- Knowledge is read on every run. It is background, not the task you give the agent.
- Files from a connected source such as Google Drive stay in the cloud and are kept in sync in a folder. A folder is read when needed, so nothing is indexed and syncing costs nothing. If you want them to be searchable, import them into a knowledge base from the folder later.
- Uploading a file to a knowledge base is billed to your wallet. The estimated cost is shown after the upload.
- Skills (reusable instructions) and Rules (what the agent must always or never do) are attached in the same area.
- You can also set an Output destination: results always appear in the run, and you can optionally save them to a knowledge base so other agents can use them.
See Knowledge for how knowledge bases and folders differ.
Memory
Memory holds facts the agent has stored, for example "the user prefers short replies", so it can use them in later runs.
- Open the Memory tab to see what the agent remembers. You can Edit a memory, Forget it, or use Show archived.
- An agent only stores something when it has the Memory capability attached under Actions, and when its instructions tell it to remember things. Without that capability the tab says the agent does not store memories.
- If a run is not watched, such as a scheduled run, a memory normally waits for your approval. Turn on Remember without asking to let the agent store what it learns in such runs. It is off by default.
Memory belongs to the agent, not to one chat. Check it now and then and forget anything that is wrong or out of date.