Persistent memory with Dakera¶

This example retrieves earlier statements from one Dakera namespace, sends them as reference data to an Instructor model, and returns a validated Pydantic response. The caller can select an exact quote from the current message to store after validation.

Connect the services¶

Run a Dakera server and install the official Python SDK. From a checkout of Instructor, set DAKERA_API_KEY, DAKERA_AGENT_ID, and OPENAI_API_KEY. The example uses the local Dakera server at http://localhost:3300 unless you set DAKERA_BASE_URL. Use a Dakera key scoped to the namespaces your application is allowed to access.

uv run --with 'dakera>=0.14,<0.15' python -m examples.persistent_memory_dakera.run

Recall and store a confirmed quote¶

Choose agent_id from the authenticated application context. Each request supplies its own namespace, current message and optional quote. Retrieved text is JSON-encoded inside a user message. The model provider receives those statements along with the current question.

from examples.persistent_memory_dakera.dakera_memory import create_with_memory
from examples.persistent_memory_dakera.run import Reply

reply = create_with_memory(
    model, memory, Reply,
    agent_id=authenticated_agent_id,
    message="I prefer dark mode. Which display setting should I use?",
    remember_quote="I prefer dark mode.",
)

Omit remember_quote for a read-only turn. For a profile or extraction task, pass your own Pydantic response model. A failed model call or validation leaves storage untouched. A confirmed quote is stored as supplied, so a reviewer can trace it back to the current input.

Use asynchronous clients¶

Use AsyncDakeraClient and an async Instructor client. The helper awaits recall, the model response and any requested write in order.

from examples.persistent_memory_dakera.dakera_memory import create_with_memory_async

reply = await create_with_memory_async(
    async_model, async_memory, Reply,
    agent_id=authenticated_agent_id,
    message="What display preference did I mention?",
)

Forget a stored statement¶

Use the ID of the record you want to remove and the same application-selected namespace. The SDK sends the scoped deletion to Dakera.

memory.forget(agent_id=authenticated_agent_id, memory_id=selected_memory_id)

The runnable example uses the public methods in Dakera Python 0.14. Example source · Dakera Python SDK · Dakera REST API