VertexAiMemoryBankService sends the model's thought parts to Memory Bank #7850

🔴 Required Information

Describe the Bug:
VertexAiMemoryBankService sends the model's thought parts (Part.thought=True) to Memory Bank as conversation content. Both add_session_to_memory and add_events_to_memory pass event.content through unchanged, on the ingest_events path and on the memories.generate path. Memory Bank then extracts memories from the model's private reasoning as if it were part of the conversation, so a guess the model made about the user can be stored as a fact about them.

An event whose only text is a thought is not filtered out either, because _is_empty_content counts any part.text as content.

The other memory services already exclude thoughts: InMemoryMemoryService and VertexAiRagMemoryService (#7473), and SqliteMemoryService (#7691). #7830 reports the same problem in FirestoreMemoryService.

Steps to Reproduce:

1. Build a session with two model events: one with parts [Part(text="the user sounds anxious, keep the reply short", thought=True), Part(text="Here is the answer.")], and one whose only part is a thought.
2. Call add_session_to_memory(session) with the agentplatform.Client mocked the way the existing unit tests do.
3. Inspect the events passed to memory_banks.ingest_events.

Expected Behavior:
The first event is sent with only Here is the answer., and the thought-only event is not sent.

Observed Behavior:
Both events are sent with their thought parts:

901 [('the user sounds anxious, keep the reply short', True), ('Here is the answer.', None)]
902 [('the user sounds anxious, keep the reply short', True)]

add_events_to_memory with a generate-only key such as revision_ttl sends the same content to memories.generate.

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at 4f8e4ae)
• Desktop OS: macOS
• Python Version (python -V): 3.14.7

Model Information:

• Are you using LiteLLM: No
• Which model is being used: any thinking model (e.g. Gemini with include_thoughts=True)

🟡 Optional Information

Additional Context:
In src/google/adk/memory/vertex_ai_memory_bank_service.py:

• _add_events_to_memory_from_events and _add_events_to_memory_via_ingest build each direct-contents event with content=event.content.
• _should_filter_out_event relies on _is_empty_content, which does not look at part.thought.

I verified this with the mocked client; I don't have a Memory Bank instance.

Minimal Reproduction Code:

import asyncio
from unittest import mock

from google.adk.events.event import Event
from google.adk.memory import VertexAiMemoryBankService
from google.adk.sessions.session import Session
from google.genai import types

THOUGHT = "the user sounds anxious, keep the reply short"


async def main():
  with mock.patch("agentplatform.Client") as client:
    aio = mock.MagicMock()
    aio.memory_banks.ingest_events = mock.AsyncMock()
    client.return_value = mock.MagicMock(aio=aio)
    service = VertexAiMemoryBankService(
        project="p", location="us-central1", agent_engine_id="123"
    )
    session = Session(
        app_name="app",
        user_id="u",
        id="s",
        last_update_time=0,
        events=[
            Event(
                id="901",
                author="agent",
                content=types.Content(
                    role="model",
                    parts=[
                        types.Part(text=THOUGHT, thought=True),
                        types.Part(text="Here is the answer."),
                    ],
                ),
            ),
            Event(
                id="902",
                author="agent",
                content=types.Content(
                    role="model",
                    parts=[types.Part(text=THOUGHT, thought=True)],
                ),
            ),
        ],
    )
    await service.add_session_to_memory(session)
    await asyncio.sleep(0)
    source = aio.memory_banks.ingest_events.call_args.kwargs[
        "direct_contents_source"
    ]
    for event in source.events:
      print(event.event_id, [(p.text, p.thought) for p in event.content.parts])


asyncio.run(main())

How often has this issue occurred?:

• Always (100%), with thinking models