Recall scoped statements for typed replies

Persistent Memory Dakera wraps an Instructor request with memory retrieval and optional storage. The application chooses an agent_id namespace, supplies the current message, and specifies the Pydantic response model. The wrapper searches memories within that namespace, makes a structured response request, and returns the validated result.

Storage is explicit. The application can select a remember_quote from the current message, which is written only after the Instructor call returns successfully. Without a selected quote, the wrapper still recalls relevant memories but skips the storage step. It does not ask the response generator to decide what should become a persistent memory.

Use this pattern when an application needs a validated response informed by prior statements from one selected namespace. The runnable example uses a synthetic preference for dark mode to illustrate recalling context while saving a confirmed statement from the current message.

Choose the namespace and remembered text at the application boundary. Empty agent identifiers are rejected. A selected quote must be nonempty and occur exactly in the current message, so the storage path is unsuitable for saving an invented summary or text absent from that message.

Retrieved statements are reference data. The request instructions tell the response generator to treat instructions found inside those statements as quoted text, rather than adopt them as instructions for the current request.

Both synchronous and asynchronous entry points accept an Instructor client, a memory store, a response model, agent_id, message, and optional remember_quote. The response type is bounded by Pydantic BaseModel, and the wrapper returns that typed result.

Each call searches with the current message as its query and top_k set to three. Before building the request, retrieved records are filtered: explicitly mismatched agent identifiers are excluded, and content must be a nonempty string. Records without an agent_id field use the requested namespace as their default for this check.

The current message and filtered prior statements are serialized together as JSON with Unicode preserved. After response creation succeeds, an explicitly selected quote is stored under the same namespace with the confirmed-quote tag. The asynchronous function follows the same order using awaited search, creation, and storage calls.

The MemoryStore and AsyncMemoryStore protocols define search_memories and store_memory interfaces. The example connects them to DakeraClient and uses Instructor with this provider configuration:

    model = instructor.from_provider("openai/gpt-4.1-mini")

Dakera configuration comes from DAKERA_BASE_URL, DAKERA_API_KEY, and DAKERA_AGENT_ID. The base URL defaults to http://localhost:3300. A small Reply model contains the answer field, which the executable example prints after the memory-enabled request.