crewai-memory-cosmosdb

An Azure Cosmos DB (NoSQL) StorageBackend for CrewAI's unified Memory — hierarchical scopes, categories, metadata filters, importance/recency, and native vector search via VectorDistance (diskANN index).

pip install crewai-memory-cosmosdb

from crewai import Crew
from crewai.memory import Memory
from crewai_memory_cosmosdb import CosmosDBMemoryBackend

backend = CosmosDBMemoryBackend(
    endpoint="https://<acct>.documents.azure.com:443/", key="<key>",
    database_name="crewai", container_name="memory", dimensions=3072,   # match your embedder
)
memory = Memory(storage=backend)
crew = Crew(agents=[...], tasks=[...], memory=memory)

Or once for every Crew(memory=True): set_memory_storage_factory(lambda spec: CosmosDBMemoryBackend(...)).

How it works

• A container partitioned by /scope, one document per record (id = record id), created with a vector embedding policy on /embedding (cosine, dimensions) and a diskANN vector index (vector_index_type="quantizedFlat"|"flat" to change). The policy is fixed at container creation — use a new container to change dimensions.
• search is one SQL query: ORDER BY VectorDistance(c.embedding, @q) with the scope subtree in WHERE; the score is the cosine similarity. categories / metadata_filter / min_score are applied on the candidates.
• Scope tree, category counts, listing, delete(older_than=...), reset are cross-partition SQL queries.
• The account needs Vector Search for NoSQL: az cosmosdb update -n <acct> -g <rg> --capabilities EnableServerless EnableNoSQLVectorSearch (list all existing capabilities; propagation can take a few minutes).

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

License

MIT