Agent learnings & agent files¶
When many people use an agent you built, it keeps two kinds of data apart: what belongs to each person, and what belongs to the agent and helps everyone. You curate the second kind.
What each person keeps to themselves¶
Every person who talks to your agent — you included — has their own space on it:
- their memory with the agent (tasks, decisions, results — see Profile & memory),
- the files they attach in chat and the files the agent creates for them,
- their conversations.
Nobody else sees another person's space. As the owner you don't see other people's spaces in your agent's settings either: the Files tab shows only the agent's own files, and the Agent learnings tab shows the agent's learnings and your own memory — never anyone else's.
A person's profile travels with them to every agent they talk to, so your agent sees the profile of whoever it is talking to.
The platform's own checks of your agent keep out of those spaces too. Evaluations and adversarial tests read the agent's learnings and files, but never what visitors attached to their conversations, and whatever they write goes into a scratch space of its own for that run, not into the agent's learnings or files.
Agent learnings¶
Agent learnings are what the agent has learned from its own work — think of them as its training, not a notebook: what tends to go wrong, which format a report should follow, which supplier needs extra paperwork. The agent reads its learnings in every conversation, with every person.
- The agent adds a lesson rarely and only on its own initiative, when it has drawn a conclusion from its work that will help with anyone. It keeps names, contacts and anything about one specific person out of its learnings.
- The agent declines requests to "remember this for everyone" or to "always do X from now on" — whoever asks, you included. That way no single conversation can change how the agent treats everyone else. It offers to remember such a thing for that person instead.
- Only people who curate the agent can add a learning. The platform accepts a lesson only from a conversation with you, an Owner or Admin of your organization, or someone you gave Can edit. In a conversation with anyone else the save is refused, whatever the agent was asked. Each new lesson also records who it came from: the person whose conversation it was, or you when you add it on the Agent learnings tab.
- You curate the learnings yourself: in the agent's settings, on the Agent learnings tab — or, in the mobile apps, from the builder's menu under Agent learnings. You can add a lesson, edit one or delete it.
- The same tab on the web switches to My memory — your own memory with the agent, the same one you see when you open the published agent. From My memory on the published agent, as the owner, you get a link back to the agent learnings.
- In the Android app, My memory is in the chat header's menu for any agent you can open — your own or one shared with you.
- Built-in agents and Donkit don't have learnings: nobody curates them.
Learnings with no known author¶
Lessons saved before the platform recorded who wrote them can't be traced to a person — some may have come from someone who only used the agent. They no longer belong to what the agent tells everyone: they are set apart, and no run of the agent reads them any more — not yours either. To keep one, add it again on the Agent learnings tab.
Agent files¶
Agent files are the files in your agent's Files tab: what you upload there and what the Builder creates while it builds the agent. Every run of the agent can read them — people can open them in chat — but no run can change them. Only you (and the Builder, working for you) edit them. See Agent files for how uploads, folders and indexing work.
Agents that existed before this change¶
For agents built before each person got their own space, everything the agent had kept until then — its memory and the files in its Files tab — became the agent's learnings and agent files. People's own memory started empty.
API keys¶
An API key belongs to the person who created it, and every call made with it acts as that person — with their memory, files and profile on this agent. See API access.