Sessions and Session Services on the Professional Agentic Architect Exam

GCP Study Hub
November 17, 2026

A session is the event history and state of one conversation. In ADK, a session service decides where that session data lives, which becomes a primary architectural decision whenever an agent runs on stateless compute or needs to survive deployments. On the Professional Agentic Architect exam, session architecture decisions usually turn on three distinctions: how each ADK session service handles persistence and operational overhead, how session storage differs from Memory Bank, and how session storage aligns with Agent Runtime, Cloud Run, and lower-level compute options.

ADK session services

In ADK, three session services represent the main persistence options, each with different durability guarantees and operational responsibilities:

  • InMemorySessionService keeps sessions in the process memory of the running application. Everything is lost on a restart, a scale-down event, or a new revision deployment.
  • DatabaseSessionService stores sessions in a database, which persists only if the underlying database itself persists across instance lifecycles.
  • VertexAiSessionService stores sessions in Agent Platform Sessions, a managed service that survives restarts, scale-down events, and new revisions with no database to run.

Two operational details are worth keeping straight when evaluating these services. First, setting one minimum instance does not make InMemorySessionService persistent, because restarts and new deployments still happen and clear process memory. Second, DatabaseSessionService does not provide durable storage if the database itself is ephemeral. A SQLite file inside a Cloud Run container vanishes along with the instance, while connecting DatabaseSessionService to a real external database means the team takes on patching, scaling, and backing up that database.

With VertexAiSessionService, Agent Platform Sessions provides managed persistence without database operations. Agent Platform Sessions is identified by an Agent Runtime instance, so a team creates an Agent Runtime instance even if the agent itself runs on Cloud Run.

Sessions versus Memory Bank

Sessions and Memory Bank serve different roles in ADK. A session holds the event history and state of one conversation. Memory Bank, which ADK reaches through VertexAiMemoryBankService, holds extracted long-term facts and memories rather than the event history of a session.

Because of that distinction, Memory Bank is not a substitute for session storage. When an agent requires both conversation history and extracted long-term memory, the design pairs a persistent session store with Memory Bank rather than using Memory Bank alone.

Pairing session storage with Agent Runtime and Cloud Run

Session storage choices connect directly to the compute platform hosting the agent. Google's architecture guidance recommends Agent Runtime for Python agents that need a fully managed experience with minimal operational overhead. Agent Runtime includes managed Agent Platform Sessions, Memory Bank for long-term memory, autoscaling, and built-in logging, monitoring, and tracing. Teams that want to focus on agent logic instead of operating underlying platform infrastructure can run Python ADK agents directly on Agent Runtime.

Agent Runtime is built for Python agents and does not support hosting custom MCP servers. When an architecture needs language flexibility such as TypeScript, serverless event-driven scaling down to zero, or hosting custom MCP clients and servers over the streamable HTTP transport, Google recommends Cloud Run for those containerized services.

Because Cloud Run services are stateless, session and memory data belong outside the container. Teams running ADK agents on Cloud Run can use VertexAiSessionService backed by Agent Platform Sessions, creating an Agent Runtime instance as its identifier so conversation history survives new revisions and scale-down events without operating a database. Firestore is another store teams use behind stateless platforms such as Cloud Run. Teams can also build and operate their own session and memory storage on Memorystore and Cloud SQL, though that adds platform work that managed options avoid.

Where GKE and managed instance groups fit

On the Professional Agentic Architect exam, lower-level compute options also appear alongside Agent Runtime and Cloud Run, and the same operational tradeoffs apply. GKE fits complex stateful requirements and workloads that need fine-grained infrastructure control, which requires a team to operate and manage it. A managed instance group leaves scaling, patching, container management, sessions, memory, and tracing entirely to the team.

When minimal operational overhead is the goal, we would lean toward Agent Runtime for managed Python agents, or Cloud Run paired with Agent Platform Sessions or Firestore when containerized language flexibility or custom MCP servers are required.

Our Professional Agentic Architect practice exams test ADK session services alongside Memory Bank and agent deployment choices, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow