Documentation
Agents and memory
What an agent knows, what it may do, and what it remembers between runs.
What an agent is
An agent is an instruction, a model, a list of allowed commands, an approval policy and its memory. It reads what a step gives it, decides what to do, calls the commands it is allowed to call, and returns a result.
The allowed commands are the agent's entire reach. If a command is not on the list, the agent cannot call it, whatever its instruction says and whatever it is asked to do.
Three kinds of memory
These are separate capabilities, switched on independently. None of them puts anything into an agent's context automatically. The agent retrieves when it decides it needs to, and that retrieval is recorded like any other action.
- Long-term facts
- The agent records durable facts and recalls them later. Useful for what a process learns over time, such as a client's standing instruction.
- Knowledge search
- The agent searches the knowledge bases attached to it. Useful for policies, manuals and reference material.
- Chat memory
- The agent remembers a conversation across runs, scoped to a workflow and a department. Off by default.
Approvals on an agent
An agent can be required to stop for a person before it acts. That can apply to everything it does, or to named commands only, so that reading is free and writing is gated. A member may add or narrow this requirement; only an organization admin can remove it entirely.
Models
Agents run on open-weight models hosted in Switzerland. The model is part of the agent's configuration and is reloaded at each step, so changing it takes effect without republishing the workflows that use the agent.