crewai-memory-firestore

A Google Firestore StorageBackend for CrewAI's unified Memory — hierarchical scopes, categories, metadata filters, importance/recency, and native vector search via find_nearest.

pip install crewai-memory-firestore

from crewai import Crew
from crewai.memory import Memory
from crewai_memory_firestore import FirestoreMemoryBackend

backend = FirestoreMemoryBackend(project_id="my-project", collection="crewai_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: FirestoreMemoryBackend(...)). Authentication uses Application Default Credentials.

How it works

• One collection, one document per record (doc id = record id) with the embedding stored as a Firestore Vector.
• search uses native find_nearest (cosine) after a scope pre-filter. A scope subtree is two queries (the exact scope, and the scope/… range) merged by similarity; score is 1 - distance. categories / metadata_filter / min_score are applied on the candidates.
• Firestore requires a composite vector index scope ASC + embedding (flat, dims). With create_index=True (default) the backend creates it through the Admin API if missing; the build is asynchronous and vector queries fail with FAILED_PRECONDITION until it is READY (a few minutes). Needs datastore.indexes.create/list, or create it yourself and pass create_index=False:

gcloud firestore indexes composite create --collection-group=crewai_memory --query-scope=COLLECTION \
  --field-config=order=ASCENDING,field-path=scope \
  --field-config=vector-config='{"dimension":"3072","flat":"{}"}',field-path=embedding

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

License

MIT