Memory

Memory#

PyRIT’s memory component allows users to track and manage a history of interactions throughout an attack scenario. This feature enables the storage, retrieval, and sharing of conversation entries, making it easier to maintain context and continuity in ongoing interactions.

To simplify memory interaction, the pyrit.memory.CentralMemory class automatically manages a shared memory instance across all components in a session. Memory must be set explicitly.

Manual Memory Setting:

At the beginning of each notebook, make sure to call:

# Import initialize_pyrit
# Import the specific constant for the MemoryDatabaseType, or provide the literal value
from pyrit.common import initialize_pyrit, IN_MEMORY, DUCK_DB, AZURE_SQL

initialize_pyrit(memory_db_type: MemoryDatabaseType, memory_instance_kwargs: Optional[Any])

The MemoryDatabaseType is a Literal with 3 options: IN_MEMORY, DUCK_DB, AZURE_SQL. (Read more below)

  • initialize_pyrit takes the MemoryDatabaseType and an argument list (memory_instance_kwargs), to initialize the shared memory instance.

Memory Database Type Options#

IN_MEMORY: In-Memory DuckDB Database

  • This option can be preferable if the user does not care about storing conversations or scores in memory beyond the current process. It is used as the default in most of the PyRIT notebooks.

  • Note: In in-memory mode, no data is persisted to disk, therefore, all data is lost when the process finishes (from DuckDB docs)

DUCK_DB: Persistent DuckDB Database

  • Interactions will be stored in a persistent DuckDBMemory instance with a location on-disk. See notebook here for more details.

AZURE_SQL: Azure SQL Database

  • For examples on setting up AzureSQLMemory, please refer to the notebook here.

  • To configure AzureSQLMemory without an extra argument list, these keys should be in your .env file:

    • AZURE_SQL_DB_CONNECTION_STRING

    • AZURE_STORAGE_ACCOUNT_DB_DATA_CONTAINER_URL