crewai-dxmt

DXMT, the open-source memory database for personal agents, as CrewAI's memory storage backend, knowledge storage and search tool.

pip install crewai-dxmt

A DXMT server has to be running: vectordb serve, or docker run -p 127.0.0.1:9999:9999 -v dxmt:/data dxmtai/dxmt.

Memory

CrewAI's unified Memory does the thinking (an LLM infers scope, categories and importance; recall blends similarity, recency and importance) and hands the storage records with their embeddings. DXMTStorageBackend implements CrewAI's StorageBackend protocol over one DXMT table, so every selection the protocol asks for -- scope prefix, categories, metadata, older_than -- is a server filter.

from crewai import Memory
from crewai_dxmt import DXMTStorageBackend

memory = Memory(storage=DXMTStorageBackend("crewai_memory", host="localhost", port=9999))
memory.remember("We decided to use PostgreSQL for the user database.")
memory.recall("What database did we choose?")

Crew(memory=True) builds its Memory without a storage argument, so to route every crew to DXMT register the backend once at startup:

import crewai_dxmt
crewai_dxmt.install(host="localhost", port=9999)

Why DXMT rather than the LanceDB default: a server process, so agent memory and crew memory, or several crews on one machine, share one store with no file lock; a write-ahead log, so kill -9 loses nothing acknowledged; and the same store is reachable over MCP, REST and the Python, JavaScript, Rust and Go clients.

Knowledge

DXMTKnowledgeStorage is a BaseKnowledgeStorage. It embeds on the server (embedding={"model": ...}, any OpenAI-compatible endpoint via base_url) or with a function (embedder=); with enable_full_text_search=True queries are hybrid (vector + BM25).

from crewai import Knowledge
from crewai_dxmt import DXMTKnowledgeStorage

storage = DXMTKnowledgeStorage(table_name="docs", embedding={"model": "text-embedding-3-small"})
knowledge = Knowledge(collection_name="docs", sources=[...], storage=storage)

install() registers a knowledge factory too, using the embedder CrewAI resolved for the collection, or knowledge_embedding={...} for server-side embedding.

Tool

from crewai import Agent
from crewai_dxmt import DXMTVectorSearchTool

tool = DXMTVectorSearchTool(table_name="docs", search_mode="hybrid", limit=5)
agent = Agent(role="Researcher", goal="...", backstory="...", tools=[tool])

The query is embedded by the table's server-side model, or by embedder=; search_mode is vector, hybrid or text.

Tests

tests/test_memory_backend.py drives CrewAI's real Memory engine end to end (remember, recall with composite scoring, scopes, categories, forget, update) with a deterministic embedder and a mocked LLM, against a live server: DXMT_HOST=127.0.0.1 DXMT_PORT=9999 pytest.