How to Handle Topic Switching in Google ADK Without Carrying Old Context

[samirsinghdba]

I'm building a chatbot using Google ADK, and I’m running into an issue with how context is maintained within a session.

If a user is having a conversation and he mentions a project_id (just as an example) early in the conversation, ADK correctly uses it for follow-up questions, which is correct, and that conversation come to an end. However, if the same user later asks a completely different question—sometimes 20–30 minutes later— and mentions project_id, but in a different context, ADK still uses that old project_id even though the new question is unrelated.

I understand that ADK uses the session_id to maintain a continuous conversation, similar to how a chat window works. But in real usage, users often switch topics within the same chat window, just like ChatGPT. Creating a brand-new session every time the user changes topics doesn’t feel practical or natural.

So my main question is:

What is the recommended way in Google ADK to detect when a user has changed topics and to clear or refresh the conversational context so the model doesn't reuse outdated values like a previous project_id? Do as a developer we need to take care of this or
does google ADK tracks it and is it recorded in any of their metadata table like events ?

Any guidance or established patterns for handling topic transitions within the same session would be very helpful.

Thanks!

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: Google ADK provides the primitives for managing context (primarily via session.state), but it does not automatically detect topic switches to clear outdated context. As a developer, you are responsible for implementing the logic to manage this. The recommended approach is to use the ContextFilterPlugin to limit conversational history and, if needed, explicitly clear the state within your tool's logic.

---

Hello! That's a great question, and a common challenge in building sophisticated chatbots.

You are correct that ADK uses a Session to maintain conversational context. [1] This context, including values stored in session.state, persists for the duration of the session, which can lead to the behavior you're observing when a user changes topics.

Currently, ADK does not have a built-in mechanism to automatically detect a topic change and clear the relevant context. This logic is the responsibility of the developer to implement based on the specific needs of the application.

Here are a few established patterns and strategies you can use in ADK to handle this:

1. Limit Conversation History with ContextFilterPlugin

This is often the simplest and most effective first step. The ContextFilterPlugin allows you to control how much of the conversation history is passed to the LLM for the next turn. By limiting the number of recent turns, you can prevent the model from "seeing" the outdated context from much earlier in the conversation.

You can configure it to only keep the last few interactions, which naturally "ages out" old information.

2. Explicitly Manage session.state

Since you can access and modify the session.state from your tools, you can build logic to clear it when appropriate. You can access the state through the CallbackContext or ToolContext provided to your tools. [2]

For example, you could:

• Create a "reset" tool: Implement a simple tool that a user can invoke with a phrase like "let's start over" or "new topic," which clears specific keys (or all keys) from session.state.
• Add logic to your existing tools: Before a tool executes, it could check if the new prompt seems to represent a new topic. This is more complex and might involve an LLM call to classify the user's intent, but it offers the most seamless user experience. If a topic change is detected, your code can then clear the old project_id from the state before proceeding.

3. Use Stateless Agents for Specific Tasks

For parts of your conversation that should never depend on history, you can configure your LlmAgent to be stateless by setting include_contents='none'. [3] This agent will only see the current user input and will not have access to any previous turns. This is useful for one-off questions but is less ideal for a continuous, flowing conversation.

To directly answer your questions:

Do as a developer we need to take care of this?

Yes, you need to implement the logic for detecting topic changes and clearing the context.

does google ADK tracks it and is it recorded in any of their metadata table like events ?

ADK tracks the history of all events and state changes within a session, but it does not automatically identify or log a "topic switch" event. You would need to build this custom logic.

I recommend starting with the ContextFilterPlugin as it's easy to implement and might solve the issue by simply limiting how far back the agent can see.

I hope this helps!

---

Footnotes:
[1] https://google.github.io/adk-docs/sessions/
[2] https://google.github.io/adk-docs/context/
[3] https://google.github.io/adk-docs/agents/llm-agents/