Troubleshooting for Memory Bank

Preview

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

This document shows you how to resolve common issues when using Vertex AI Agent Engine Memory Bank.

No memories were generated

The memory generation process includes a crucial step: determining if information in the source content is meaningful enough to persist. An empty response indicates the process ran successfully but found no information that met the criteria for being saved. If you were expecting memories to be generated, this guide can help you identify a potential error or misconfiguration.

To troubleshoot why memories are not being generated, follow these steps in order:

Check if memory generation was triggered

First, confirm that the memory generation process was actually initiated. Memory generation is initiated by invoking GenerateMemories (client.agent_engines.memories.generate(...)).

If you're using ADK's VertexAiMemoryBankService, memory generation is not automatically triggered. You must ensure your agent or application has explicitly called the add_session_to_memory method to trigger the process.

add_session_to_memory takes a Session object as input and uses the session's events as the data source for memory generation. It only calls your Memory Bank instance if there are events populated in the session object. If your ADK application or agent is invoking add_session_to_memory but memory generation was not triggered, the Session object's events may not be populated. This is possible even if you have interacted with the session, especially if you're using adk.Runner. To address this issue, fetch the session and its events to the environment where you're invoking add_session_to_memory:

session = await session_service.get_session(
    app_name=app_name,
    user_id=user_id,
    session_id=session.id
)
# Confirm that events are populated.
print(session.events)
memory_service.add_session_to_memory(session)

Verify that the memory generation LRO is complete

Memory generation is a long-running operation (LRO) and can take a few seconds to complete. The exact latency depends on the length of the input conversation and the complexity of the information being processed.

When using the Agent Engine SDK, memory generation is a blocking operation by default. So, client.generate_memories(...) blocks your code's execution until the memory generation LRO is complete.

When using ADK's VertexAiMemoryBankService, add_session_to_memory is a non-blocking operation. It only triggers memory generation and does not wait for the LRO to complete.

Look for errors in the operation response

The LRO response may an error message similar that indicates that memory generation was unsuccessful. For example:

RuntimeError: Failed to generate memory: {'code': 3, 'message': 'Failed to extract memories: Please use a valid role: user, model.'}

Common errors include:

• Resource exhausted errors for Gemini when you use pay-as-you-go. With dynamic shared quota (DSQ), there are no predefined quota limits on your usage. To help ensure high availability for Memory Bank and to get predictable service levels for your production workloads, see Provisioned Throughput.

• Invalid source data, like using roles other than model and user in your Content.

Determine if the conversation was meaningful

If the process was triggered, completed successfully, and produced no errors, it's likely that Memory Bank determined that no information in the source conversation was meaningful enough to persist.

Memory Bank uses "memory topics" to identify what information is meaningful. If the content of your conversation doesn't align with any configured topics, no memories are generated.

If you believe information should have been persisted, you can customize your Memory Bank instance's configuration to better align with your expectations.

Configuring memory topics lets you define what information should be persisted. Configuring few-shot examples helps your Memory Bank instance adapt to your expectations by teaching it the nuances in what information should be persisted and with what phrasing. You can think of customizing your Memory Bank in two steps: Telling and Showing. Memory Topics tell Memory Bank what information to persist. Few-shots show Memory Bank what kind of information should result in a specific memory, helping it learn the patterns, nuance, and phrasing that you expect it to understand.