crewai-memory-core

Shared core for building CrewAI unified-memory StorageBackend implementations. CrewAI 1.10+ has one Memory engine (LLM analysis, consolidation, scope inference, composite scoring) over a pluggable storage protocol; this core implements the whole protocol — save / search / delete / update / get_record / list_records / get_scope_info / list_scopes / list_categories / count / reset / touch_records and the async variants — so a backend only supplies five primitives:

from crewai_memory_core import MemoryBackend

class MyBackend(MemoryBackend):
    def _put(self, rows): ...                    # upsert rows by id
    def _get(self, record_id): ...               # -> row | None
    def _delete_ids(self, ids): ...              # -> int deleted
    def _scan(self, scope_prefix): ...           # every row in the scope subtree
    # optional — native ANN; default ranks _scan rows by cosine in Python
    def _vector_search(self, vector, scope_prefix, limit): ...   # -> [(row, score), ...]

A row is a dict of the MemoryRecord fields plus embedding. Scope paths are hierarchical (/company/team/user); a prefix matches the scope itself and everything under it, never a sibling that merely shares characters (/a does not match /ab).

Use a backend with CrewAI:

from crewai import Crew
from crewai.memory import Memory

memory = Memory(storage=MyBackend(...))
crew = Crew(agents=[...], tasks=[...], memory=memory)     # or set_memory_storage_factory(...) once at startup

crewai_memory_core.testing ships FakeEmbedder (deterministic, no network — pass as Memory(embedder=...)) and InMemoryBackend; crewai_memory_core.contract is an importable test suite every provider runs, including an end-to-end pass through CrewAI's real Memory engine with zero LLM calls.

Concrete backends: crewai-memory-dynamodb, crewai-memory-postgres, crewai-memory-cosmosdb, crewai-memory-firestore.

Docs: https://skamalj.github.io/agentstate-reducer/

License

MIT