How long can a live agent session remain open for?

[enxilium]

I have a streaming tool that needs to continuously monitor something for long periods of time, and I was wondering how long the session could remain open for without manual user input as well as whether the time it remains open for counts towards the free tier quota for Gemini Live.

Thank you!

[adk-bot]

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

TLDR: There is no specific documentation on a fixed session timeout for a live agent. Sessions are designed to be resumable if they close. The free tier, Vertex AI Express Mode, has limitations and is valid for 90 days.

---

Hello! I can provide some information regarding your questions about live agent sessions.

While there is no explicit documentation stating a specific timeout for how long a live agent session can remain open without user input, the system is designed with session management that allows for sessions to be resumed. If a connection is closed, the client can be configured to attempt to reconnect. For instance, some examples show a client attempting to reconnect after 5 seconds [1]. The underlying design indicates that if a session times out, it will close without raising an error [2].

Regarding the free tier, Google offers Vertex AI Express Mode, which provides free access to services like VertexAiSessionService and VertexAiMemoryBankService. However, this free tier comes with certain limitations. Projects in Express Mode are valid for only 90 days, and there are restricted quotas on the available services. To remove these limitations, you would need to add a billing account to your project [3]. The documentation does not specify how open session time is measured against the free tier quota.

[1] https://github.com/google/adk-docs/blob/main/docs/streaming/custom-streaming-ws.md
[2] https://github.com/google/adk-python/blob/main/src/google/adk/flows/llm_flows/base_llm_flow.py
[3] https://github.com/google/adk-docs/blob/main/docs/sessions/express-mode.md

[boyangsvl]

This is depending on the model you choose. For Gemini, audio-only sessions are limited to 12 minutes. Learn more here: https://ai.google.dev/gemini-api/docs/live-session#maximum-session-duration

[enxilium]

Does the ADK support using context window compression and session resumption to extend the active session for an unlimited amount of time? It's crucial the streaming tool remains active at all times, listening for events.

Thank you!

[surajksharma07]

Following up on your last question enxilium. RunConfig now exposes both session_resumption and context_window_compression for live sessions so you can use those to keep a streaming session going past the per-connection cap. Session resumption has been around since July and compression landed a bit after this thread. They extend/resume rather than remove the underlying model-side session limit but should cover your continuous-monitoring use case. Closing this out since it looks addressed, feel free to reopen or comment if you're still running into it.