Memory 3: Service Layer

Create a package resembling:

src/server/app/memory/
|-- adapters.py
|-- config.py
|-- registry.py
|-- schemas.py
+-- service.py

Responsibilities:

• Resolve the configured database pool and model credentials.
• Translate the repository's LiteLlmModelSpec and embedding configuration into SDK Llm and Embedder instances.
• Cache components by database, model, and configuration identity, not by user.
• Invalidate the cache after relevant configuration or credential changes.
• Close SDK components during application shutdown.
• Require a MemoryPrincipal on every scoped call.
• Always search with exact user_id matching.
• Return project-owned Pydantic DTOs rather than exposing SDK objects directly.
• Map SDK exceptions into stable API error categories.
• Never log prompts, stored content, search queries, metadata, user IDs, or thread IDs.

Recommended service methods:

• create_thread
• get_thread
• append_turn
• get_messages
• add_memory
• search_memories
• update_memory
• delete_memory
• delete_thread
• get_summary
• get_context_card

Include request or turn IDs and deterministic message IDs for retry safety. This matters for streaming reconnects and multi-replica deployments.

Start with the proposed src/server/app/memory/ package layout and review the responsibilities for adapters.py, config.py, registry.py, schemas.py, and service.py. Use the listed service methods as the scope, including principal checks, exact user matching, caching, invalidation, shutdown, DTOs, error mapping, and retry-safe IDs. Done means the service satisfies these requirements without exposing sensitive data in logs.