agent-framework-dxmt

DXMT, the open-source memory database for personal agents, for Microsoft Agent Framework: a vector store with keyword-hybrid search, and memory as a context provider.

pip install agent-framework-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). Connection settings are DXMT_URL (default http://localhost:9999) and DXMT_API_KEY, or the url / api_key arguments, resolved through Agent Framework's load_settings like every other connector.

Vector store

DXMTCollection and DXMTStore implement BaseVectorCollection / BaseVectorSearch and BaseVectorStore. A collection is one DXMT table; the record's single vector field is the row's vector and its data fields are the payload.

from dataclasses import dataclass
from typing import Annotated

from agent_framework import Filter, VectorStoreField, vectorstoremodel
from agent_framework_dxmt import DXMTStore


@vectorstoremodel
@dataclass
class Document:
    id: Annotated[str, VectorStoreField("key")]
    title: Annotated[str, VectorStoreField("data", is_full_text_indexed=True)]
    year: Annotated[int, VectorStoreField("data", is_indexed=True)]
    embedding: Annotated[list[float] | None, VectorStoreField("vector", dimensions=3)] = None


async with DXMTStore() as store:
    collection = store.get_collection(Document, collection_name="documents")
    await collection.ensure_collection_exists()
    await collection.upsert([Document("1", "Hello DXMT", 2026, [1.0, 0.0, 0.0])], generate_vectors=False)
    results = await collection.search(vector=[1.0, 0.0, 0.0], filter=Filter("year", "gte", 2020), top=3)
    async for item in results:
        print(item["record"].title, item["score"])

What is specific to DXMT:

• Keyword-hybrid search. A data field marked is_full_text_indexed gets the table's BM25 index, and collection.search("hello", search_type="keyword_hybrid", additional_property_name="title") is the server's hybrid search (vector and keywords fused on the server). Most in-tree connectors declare {"vector"} only.
• Server-side embedding. DXMTCollection(..., server_embedding={"model": "text-embedding-3-small"}, server_embedding_field="title") has the server embed records whose vector is None and queries given as text, so no embedding client runs in the agent's process.
• Filters. Every portable operator (eq, ne, ranges, between, in, not_in, is_null, is_not_null, exists, contains, contains_any, contains_all, starts_with, ends_with, contains_text) and and / or / not groups, translated exactly to the server's flat filter; a shape the server has no flat form for ((a and b) or c) raises NotImplementedError rather than being approximated.
• Scores are similarities (1.0 is identical under cosine); score_threshold is in those units. One vector field per collection; cosine_similarity (default), dot_prod, euclidean_distance.

Memory as a context provider

DXMTContextProvider recalls memories before a run (by meaning and keyword) and remembers the exchange after it, using DXMT's own memory API. Scope is tags: memories are written under user_id / agent_id / application_id and recalled by the separate search_* ids, the split the mem0 provider uses so a memory written under a shared agent is not read back for an unrelated user by accident.

from agent_framework import Agent
from agent_framework_dxmt import DXMTContextProvider

memory = DXMTContextProvider(user_id="alice", search_user_id="alice")
agent = Agent(chat_client, instructions="...", context_providers=[memory])
await agent.run("I prefer answers in Chinese.")
await agent.run("How should you answer me?")   # the preference is recalled into the context

memory.remember(text, kind=..., supersedes=[...]), memory.recall(query) and memory.forget(ids) expose the three operations directly. A memory said twice is kept once; a new fact can supersede an old one, which is kept as history and left out of recall. The same store is what every other agent on the machine sees over MCP and REST.

Tests

pytest                                  # unit tests, client mocked
DXMT_TEST_URL=http://127.0.0.1:9999 pytest -m integration   # against a server