ADK Plugin: Memory

Overview

The ADK Memory services provide ADK-compatible memory implementations that integrate with FloTorch’s memory infrastructure.

Exports:

• FlotorchMemoryService – Traditional memory search using FloTorch Memory
• FlotorchADKVectorMemoryService – Vector-based memory search using FloTorch VectorStore

---

Setup

from flotorch.adk.memory import FlotorchMemoryService, FlotorchADKVectorMemoryService

API_KEY = "<your_api_key>"
BASE_URL = "https://gateway.flotorch.cloud"
PROVIDER = "<flotorch_memory_provider>"
VECTORSTORE_ID = "<your_vectorstore_id>"

---

Memory Services

FlotorchMemoryService

FlotorchMemoryService(
  name: str,
  base_url: str | None = None,
  api_key: str | None = None,
)

Uses FloTorch Memory for traditional memory operations.

add_session_to_memory(…) -> None

async def add_session_to_memory(self, session: Session) -> None

Adds an ADK session to FloTorch memory with automatic role mapping and content extraction.

search_memory(…) -> SearchMemoryResponse

async def search_memory(
  self,
  *,
  app_name: str,
  user_id: str,
  query: str
) -> SearchMemoryResponse

Searches memory using FloTorch Memory with user/app filtering.

---

FlotorchADKVectorMemoryService

Constructor:

FlotorchADKVectorMemoryService(
  api_key: str,
  base_url: str,
  vectorstore_id: str | None = None,
)

Uses FloTorch VectorStore for semantic memory search.

search_memory(…) -> SearchMemoryResponse

async def search_memory(
  self,
  query: str,
  knn: int | None = None,
  **kwargs
) -> SearchMemoryResponse

Performs vector-based semantic search with optional KNN parameter.

---

Features

Role Mapping

Automatically maps ADK roles to FloTorch expected roles:

• model/assistant/agent/bot/ai → assistant
• user/human/person → user
• system → system
• tool → tool

Content Extraction

Robustly extracts text content from ADK Content and Part objects.

Error Handling

• Graceful fallbacks on content extraction failures
• Empty responses on search errors instead of exceptions

---

Usage Example

from google.adk.runner import Runner

# Traditional memory
memory_service = FlotorchMemoryService(
    name="mem0",
    base_url=BASE_URL,
    api_key=API_KEY
)

# Vector memory
vector_memory = FlotorchADKVectorMemoryService(
    api_key=API_KEY,
    base_url=BASE_URL,
    vectorstore_id=VECTORSTORE_ID
)

# Use with Runner
runner = Runner(
    agent=agent,
    memory_service=memory_service  # or vector_memory
)

---

Notes

• Memory services are designed to work seamlessly with ADK’s Runner
• Vector memory requires a configured vector store in FloTorch
• Both services handle ADK session and event objects automatically