VertexAiMemoryBankService

src/google/adk/memory/vertex_ai_memory_bank_service.py の Python 実装を、Java の BaseMemoryService およびメモリサービスを受け取る Runner コンストラクターと比較します。LlmAgent がメモリをどのように扱うかを追跡し、続いて、言及されている search_memory ツールと auto_save_to_memory_callback のエントリーポイントを調べます。これらの要素がすでに統合されているか、または Java にどの機能が不足しているかを文書化できれば完了です。

I’m interested in using VertexAiMemoryBankService for long-term memory when working with an agent through ADK. I see that the VertexAiMemoryBankService https://github.com/google/adk-python/blob/632bf8b0bcf18ff4e4505e4e5f4c626510f366a2/src/google/adk/memory/vertex_ai_memory_bank_service.py#L38 is implemented in the Python version of ADK, but it’s not yet available in the Java version.

By analogy with the Python version, I can implement my own Java version of VertexAiMemoryBankService (implements BaseMemoryService). I can also create a custom version of VertexAiRunner based on com.google.adk.runner.Runner#Runner(com.google.adk.agents.BaseAgent, java.lang.String, com.google.adk.artifacts.BaseArtifactService, com.google.adk.sessions.BaseSessionService, com.google.adk.memory.BaseMemoryService).

My question is the following: is it enough to simply implement my own version of VertexAiMemoryBankService and pass it into the Runner constructor, since the Java ADK already has the functionality for working with long-term memory implemented at this point? In other words, does the Java ADK already know how to add and retrieve the necessary information from memory by itself for LlmAgent? Or, apart from implementing my own VertexAiMemoryBankService and passing it to the Runner constructor, would I need to do anything else to make it work? What is the current state of this memory-related functionality as of now?

It seems to me that, in addition to providing the VertexAiMemoryBankService to the agent, I also need to provide the search_memory tool (so that the agent can search the memory by itself) and add a callback for writing to the Memory Bank after each user prompt and LLM response using after_agent_callback=[auto_save_to_memory_callback]? Both, and both the search_memory tool and the auto_save_to_memory_callback should internally use the corresponding methods from VertexAiMemoryBankService, right?