mcp-agent-mongodb 0.1.1

MongoDB Atlas-backed persistent agent memory for mcp-agent (LastMile AI): a drop-in Memory backend with Atlas Vector Search recall.

mcp-agent-mongodb

⚠️ ALPHA — NOT AN OFFICIAL MONGODB PRODUCT. This integration is in Alpha and is not a supported or official MongoDB product. Use at your own risk.

MongoDB Atlas–backed persistent memory for mcp-agent (LastMile AI) — a drop-in replacement for the framework's in-process SimpleMemory.

• MongoMemory — implements mcp-agent's Memory contract (extend / set / append / get / clear), so you assign it directly to llm.history. Each message is one MongoDB document, scoped by session_id and ordered by an append seq, so an agent's conversation history survives restarts and can be shared across processes.
• Semantic recall — optional long-term recall over past messages via Atlas Vector Search ($vectorSearch). Embedding source-agnostic: bring your own query vector (default) or enable Atlas Automated Embedding (server-side embeddings, no client code).

Capabilities

Capability	How
Persistent agent memory (MS)	MongoMemory as a drop-in llm.history
Semantic recall (VS)	recall_semantic() over Atlas $vectorSearch, session-prefiltered
Survives restarts / multi-process	history keyed by session_id, stored in MongoDB
TTL expiry	optional TTL index on ts

Why

mcp-agent's every AugmentedLLM keeps conversation history in a Memory object; the default SimpleMemory holds it in RAM and loses it when the process ends. MongoMemory is a database-backed Memory: drop it in and the agent's history is durable, queryable, and shareable — with optional semantic recall over everything it has seen.

Architecture

AugmentedLLM.history  ──►  MongoMemory(connection_string, session_id=…)
   append/extend/set/get/clear        │
                                       ▼
                          MongoDB / Atlas  "memory" collection
                          { session_id, seq, role, message, content, embedding?, ts }
                                       │
                       recall_semantic ▼  (optional)
                          Atlas Vector Search  $vectorSearch (session-prefiltered)

Install

pip install mcp-agent-mongodb

Quick start (drop-in persistent history)

from google.genai import types
from mcp_agent.agents.agent import Agent
from mcp_agent.workflows.llm.augmented_llm_google import GoogleAugmentedLLM
from mcp_agent_mongodb import MongoMemory

async with agent:
    llm = await agent.attach_llm(GoogleAugmentedLLM)
    # Swap the in-process history for a MongoDB-backed one:
    llm.history = MongoMemory(
        "mongodb+srv://...",
        session_id="user-123",
        message_model=types.Content,   # rehydrate provider message objects on read
    )
    # ...subsequent runs with the same session_id reload this history from MongoDB.

MongoMemory matches the Memory contract exactly: extend(messages), set(messages), append(message), get() -> list, clear().

Options

Arg	Default	Purpose
connection_string	—	MongoDB / Atlas URI (required)
session_id	—	Conversation/agent scope (required); every read/write is filtered by it
database_name	mcp_agent	Database name
collection_name	memory	Collection name
message_model	None	Optional pydantic model to rehydrate stored messages on get()
vector_search_index	idx_agent_memory	Atlas Vector Search index name
auto_embed	False	Enable Atlas Automated Embedding (recall by query text)
auto_embed_model	voyage-4	Voyage model used by Automated Embedding
ttl_seconds	None	If set, TTL index on ts auto-expires idle conversations

Document shape

{
  "session_id": "user-123",
  "seq": 7,
  "role": "user",
  "message": { /* the serialized message, returned verbatim from get() */ },
  "content": "summarize the Q3 report",
  "embedding": [ /* 1024 floats — bring-your-own-vector path only */ ],
  "ts": { "$date": "..." }
}

Semantic recall (Atlas Vector Search)

The package never calls an embedding provider itself — choose one of two first-class paths:

1. Bring your own vector (default).

mem = MongoMemory("mongodb+srv://...", session_id="user-123")
mem.ensure_vector_index(num_dimensions=1024)         # one-time, on Atlas

mem.append({"role": "user", "content": "I prefer window seats."},
           embedding=my_provider.embed("I prefer window seats."))

hits = mem.recall_semantic(query_vector=my_provider.embed("seating?"), k=5)

2. Atlas Automated Embedding. Atlas embeds server-side (no client embedding code):

mem = MongoMemory("mongodb+srv://...", session_id="user-123", auto_embed=True)
mem.ensure_vector_index()                            # builds an `autoEmbed` index
mem.append({"role": "user", "content": "I prefer window seats."})
hits = mem.recall_semantic(query="seating preferences", k=5)

There is no silent fallback: if you neither pass a query_vector nor enable auto_embed, recall_semantic raises ValueError.

MCP + MongoDB synergy

mcp-agent is built on the Model Context Protocol. Pair this memory backend with mongodb-partners/memory-mcp to make Atlas both the agent memory backend (this package) and an MCP memory server your agents can call as a tool.

Demos

• demo/memory_demo.py — persistence across two simulated processes + Atlas Vector Search recall (bring-your-own Voyage vectors; MEMORY_MODE=auto for Automated Embedding).
• demo/agent_demo.py — a real Gemini mcp-agent whose llm.history is a MongoMemory; Session 2 (brand-new app/agent/LLM) answers using history reloaded from Atlas.

pip install -e ".[dev]" "mcp-agent" "google-genai" voyageai
# demo/.env: ATLAS_URI, VOYAGE_API_KEY, GEMINI_API_KEY
python demo/memory_demo.py                  # bring-your-own vectors
MEMORY_MODE=auto python demo/memory_demo.py # Atlas Automated Embedding
python demo/agent_demo.py                   # Gemini agent, cross-session memory

Why MongoDB

One database for agent state: durable conversation history, semantic recall via Atlas Vector Search, TTL lifecycle, and flexible documents for arbitrary provider message shapes — no separate vector store to operate.

Conventions

• The package owns its MongoClient (built from a connection string). Connection appName = devrel-integ-mcp-agent-python and the mcp-agent-mongodb driver-info handshake are always set and non-overridable (server-side attribution).
• Embeddings use Voyage AI 3.5 (voyage-3.5, 1024-dim) on the bring-your-own path.

Tests

pip install -e ".[dev]"
pytest -q          # 17 tests, mongomock — no infra required

Resources

• mcp-agent: https://github.com/lastmile-ai/mcp-agent · docs: https://docs.mcp-agent.com
• MongoDB Atlas Vector Search: https://www.mongodb.com/docs/atlas/atlas-vector-search/
• Voyage AI embeddings: https://docs.voyageai.com

License

MIT