pydantic-ai-cosmosdb-memory

An Azure Cosmos DB (NoSQL) MemoryStore for the PydanticAI harness Memory capability — the per-user Markdown notebook the harness injects before every model call and lets the model write through write_memory, on a shared, durable container.

pip install pydantic-ai-cosmosdb-memory

from pydantic_ai import Agent
from pydantic_ai_harness.memory import Memory
from pydantic_ai_cosmosdb_memory import CosmosDBMemoryStore

store = CosmosDBMemoryStore(endpoint="https://<acct>.documents.azure.com:443/", key="<key>",
                            database_name="agents", container_name="memory")
agent = Agent("anthropic:claude-sonnet-5", deps_type=Deps,
              capabilities=[Memory(store=store, namespace=lambda ctx: ctx.deps.user_id)])

How it works

• One container partitioned by /id: a document per memory file (kind: "file", path, content, version, operation_id) and per operation receipt (kind: "op"). Database and container are auto-created under key-based auth.
• Compare-and-swap: create uses create_item (fails if the id exists); replace and delete re-read the document, check the harness version, then write with If-Match on the Cosmos ETag, so a concurrent change between read and write is rejected. A stale version raises MemoryConflictError.
• Operation receipts are reserved with create_item, completed after the mutation, dropped if it fails — durable-execution replays return the original result; a reused id with different arguments raises MemoryOperationConflictError.
• search is the harness's bounded lexical search (SearchableMemoryStore); list_paths is a STARTSWITH query on path.

Docs: https://skamalj.github.io/agentstate-reducer/ · part of pydantic-ai-memory

License

MIT