Sign in to continue learning
Building Live Voice Agents with Google’s ADK
Intermediate
Topics
Agents
In the previous lesson, we saw how to create an agent, add a built-in tool and have live conversations. But we didn't do any of the heavy lifting. ADK Web hides away most of the complexities like session and memory management. In this lesson, let's take a quick detour from the hands-on building and understand what these services are. Let's get started. In the previous lesson, we created a basic agent with a built-in tool. The ADK Web interface handles much of the underlying complexity for us. However, as you build more sophisticated agents, you will need finer control of core services like Session, State, and Memory. And that is exactly what we're going to be discussing in this lesson. Let's dive straight in. Think of Session as the container for a single conversation thread. It begins when a user starts interacting with your agent and in many cases ends when that interaction stops. For instance, if you're running an agent in a notebook, the session terminates when you stop the runtime. The session object acts as a comprehensive data log. for the conversation, capturing state, events, and metadata. ADK manages sessions through a SessionService, and the default is an InMemorySessionService, which is temporary and does not persist after the conversation ends. However, you can also configure a persistent session service to store and retrieve session data for later use. And what you see here is a very simple example of what the session might contain. It has a state, it has the events and an app name. which is basically a data dump of all the conversations, of all the logs that goes into your agent. Next is the state. So state is a session's short-term scratchpad. It is basically a dictionary of key value pairs that holds the current context of the conversation. You can write any information to the state and it will be accessible to all the tools and agents within that session. This makes it a powerful mechanism for communication. For example, if you're using multiple tools or multiple agents that are involved in the same task, they can also share information by reading and writing to the session state. And the third one is your memory service. Memory is very different from Session in that Session is actually your data dump of everything that happens in an agent, but Memory is more long-term. It provides your agent with the long-term persistence. So remember how we saw the JSON for a session? It basically holds a ton of information, but not everything is actually useful for the agent to remember long-term. So memory is a compacted version of that. information the agent needs to remember for long term. For example, if you tell your agent that your favorite coffee order is an oat milk latte, you would likely want it to remember for a long time and not forget it again tomorrow. And that is exactly what memory provides you with. In most cases, in memory services, most of the memory services uses an LLM to actually go through the session data and extract key valuable pieces of information that needs to be remembered long term and then saves it. And ADK offers different memory service implementations including an in-memory service and an integration with the Vertex AI memory bank. We'll explore the memory bank in a later session in this course. And with that, we've covered the theoretical foundations of session, state, and memory in ADK.
