FirestoreMemoryService indexes, stores and returns the model's thought parts #7830

🔴 Required Information

Describe the Bug:
FirestoreMemoryService treats the model's thought parts (Part.thought=True) as conversation text in three places:

1. It builds an event's search keywords from every part, including thoughts, so words that only appear in the model's hidden reasoning make the memory match a query.
2. It stores the full content, thoughts included.
3. search_memory returns that content as-is, so load_memory / PreloadMemoryTool hand the model its own earlier reasoning back as a "memory".

An event whose only text is a thought still gets a memory document.

The other memory services already exclude thoughts. InMemoryMemoryService skips them when scoring and drops them from results, and VertexAiRagMemoryService skips them at ingest (#7473). SqliteMemoryService drops them from search results (#7691).

Steps to Reproduce:

1. add_session_to_memory with a session holding a model event whose parts are [Part(text="weighing the hidden escalation", thought=True), Part.from_text(text="quick brown fox")].
2. Look at the document written by batch.set, then search_memory(query="fox").

Expected Behavior:
Keywords come from quick brown fox only, the stored and returned content hold only the visible part, and a thought-only event is not stored.

Observed Behavior:
The keywords include weighing, hidden and escalation. The stored and returned content include the thought part, and a thought-only event is stored.

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at 75fb788)
• Desktop OS: macOS
• Python Version (python -V): 3.11.16

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/integrations/firestore/firestore_memory_service.py:

• _write_memories does text = " ".join([part.text for part in event.content.parts if part.text]) and stores event.content.model_dump(...).
• _search_by_keyword validates and returns the stored content unchanged.

I verified this with the mocked Firestore client the existing tests use; I don't have a Firestore instance.

How often has this issue occurred?:

• Always (100%), with thinking models