Self-hosted Sandbox

A self-hosted sandbox is the room where an AI agent actually does its work, built directly inside the developer's own server rather than a service provider's cloud.

Think of it like this: instead of renting outside office space, a company builds an isolated workroom inside its own building and only lets its robot employee (the agent) work there. Because the room is isolated, even if the robot accidentally deletes the wrong files or runs a strange command, the damage doesn't spread beyond that room to the rest of the system. At the same time, because the room sits inside the company's own building, the data never leaves for outside servers and stays within the organization's own infrastructure — that's what sets it apart from the cloud approach.

Recently, a feature has started to appear that lets this room save what the agent produces so it can be remembered for the next task. Even though the room lives on your own server, the experience built up inside it can now carry over and keep being used.

In articles, you'll see phrasing like "work done in a self-hosted sandbox can now be saved to memory." Here, "sandbox" doesn't refer to a virtual space for security testing — it refers to the actual execution environment where the agent carries out real work. The "self-hosted" part means that execution environment runs on the developer's own server, not on the service provider's.

1. Try adding a memory_store type resource to the resources array when creating a session.
2. Example code: session = client.beta.sessions.create( agent=agent.id, environment_id=env.id, resources=[{"type": "memory_store"}] )
3. Work results produced in a session created this way can then be referenced in the next session running on your own server as well.