memory_store

A memory_store is a place where an AI agent writes down the results it produces or the information it learns while working. Think of it like a handover note left by a part-time worker at the end of a shift. Without the note, the next person has to figure everything out from scratch each time; with it, they can pick up right where things left off.

Until now, even when an agent worked inside an isolated execution environment run directly on a developer's own server, the results often disappeared once that work session ended. Attaching a memory_store means that even in this kind of self-hosted environment, previous work stays on record, so the next session can refer back to it and continue.

In short, a memory_store is the component that gives an agent 'long-term memory' on top of what used to be only 'short-term memory.' Since it works like an option you specify when creating a new session, teams running agents on their own infrastructure rather than in the cloud can also maintain continuity across work sessions.

In articles, it shows up in phrasing like "you attach a resource of type memory_store when creating a session." A common misunderstanding is that this is not a chat-history feature that saves entire conversations. A memory_store is closer to a developer-facing setting: it lets the results or reference data an agent produces during work persist as a resource attached to the session, so the next session can pick it up and continue.

If you have access to a hands-on development environment, try adding a resource list to a request that creates an agent session, and include one item of type memory_store in it. Then, when creating a new session in the same execution environment, reconnect the same store and check whether the agent continues to reference the information left behind from the previous session.