ADK Long-Term Memory
While Session State manages short-term conversational context, the Memory Service in ADK provides agents with long-term recall across multiple sessions. It acts as a searchable archive of past interactions and external knowledge.
1. Core Concepts
- Long-Term Knowledge: Information that persists even after a session is closed.
- Semantic Search: The ability to retrieve past context based on meaning (using embeddings) rather than just keyword matching.
- Memory Integration: Like other services, a
MemoryServiceis provided to theRunnerat initialization.
2. Service Implementations
InMemoryMemoryService
Stores memory in the application's RAM.
- Search Method: Basic keyword matching.
- Use Case: Prototyping and simple local testing.
- Persistence: Data is lost on restart.
VertexAiMemoryBankService
A production-grade service using Google Cloud's Vertex AI.
- Search Method: Semantic (vector) search using embeddings.
- Use Case: Applications requiring robust, persistent, and "intelligent" recall across thousands of sessions.
DatabaseMemoryService
A community-supported persistent memory service for python (backed by SQLAlchemy/Postgres).
3. The Memory Lifecycle
Saving to Memory
When a session concludes, it can be "committed" to long-term memory.
# Python Example
await memory_service.add_session_to_memory(completed_session)Retrieving from Memory
Agents (or tools) can search memory during an active turn.
- Input: A query string.
- Output: A list of
MemoryResultobjects, containing relevant event snippets or session data.
4. State vs. Memory
| Feature | Session State | Memory Service |
|---|---|---|
| Duration | Single Session | Cross-Session |
| Scope | Current conversation | Archive of all past conversations |
| Access | Direct Key/Value (context.state) |
Searchable (memory_service.search) |
| Storage | Volatile (Session Service) | Persistent (Memory Service) |
*Source: lit-adk-documentation*