agency-swarm-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 persistence and vector memory for VRSEN Agency Swarm.

• MongoThreadStore — drop-in load_threads_callback / save_threads_callback for the Agency class: persist entire conversations to MongoDB and restore them across restarts.
• MongoMemoryStore (new in 0.1.1) — semantic / episodic long-term memory with MongoDB Vector Search recall. Embedding source-agnostic: bring your own query vector (default) or enable Atlas Automated Embedding (server-side embeddings, no client code).

Why

Agency Swarm persists conversations through two Agency hooks but ships no database backend (only a file-based example). MongoThreadStore is that backend: one document per chat_id, idempotent upserts, optional TTL expiry — backed by MongoDB or Atlas.

Install

pip install agency-swarm-mongodb

Usage

from agency_swarm import Agency, Agent
from agency_swarm_mongodb import MongoThreadStore

store = MongoThreadStore("mongodb+srv://...", database_name="agency_swarm")
load_cb, save_cb = store.as_callbacks("user-123")   # chat_id captured in the closures

agency = Agency(
    Agent(name="Assistant", instructions="You are helpful."),
    load_threads_callback=load_cb,
    save_threads_callback=save_cb,
)

The store matches the Agency Swarm callback signatures exactly: load_threads_callback() -> list[dict] and save_threads_callback(messages: list[dict]) -> None.

Options

Arg	Default	Purpose
connection_string	—	MongoDB / Atlas URI (required unless client given)
database_name	agency_swarm	Database name
collection_name	threads	Collection name
ttl_seconds	None	If set, TTL index on updated_at auto-expires idle chats
client	None	Bring your own MongoClient (then appName is not overwritten)

Document shape

{
  "_id": "user-123",          // chat_id
  "messages": [ /* full flat list, exactly as Agency Swarm emits */ ],
  "message_count": 12,
  "updated_at": { "$date": "..." }
}

Vector memory (MongoMemoryStore, new in 0.1.1)

Semantic / episodic long-term memory with MongoDB Vector Search recall. The package never calls an embedding provider itself — you choose one of two first-class paths:

1. Bring your own vector (default). Embed with whatever provider you already use (OpenAI, Voyage SDK, Cohere, your agency's embedding client, …) and pass the vectors:

from agency_swarm_mongodb import MongoMemoryStore

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

mem.add_memory("user-123", "user", "Prefers window seats.",
               kind="semantic", embedding=my_provider.embed(text))

hits = mem.recall_semantic("user-123",
                           query_vector=my_provider.embed("seating?"), k=5)

2. Atlas Automated Embedding. Atlas generates embeddings server-side (no client embedding code); recall sends query text:

mem = MongoMemoryStore("mongodb+srv://...", auto_embed=True)   # default model: voyage-4
mem.ensure_vector_index()                            # builds an `autoEmbed` index

mem.add_memory("user-123", "user", "Prefers window seats.", kind="semantic")
hits = mem.recall_semantic("user-123", query="seating preferences", k=5)

There is no silent fallback: if you neither pass a query_vector nor enable auto_embed, recall raises ValueError. Episodic recency recall needs no vectors:

mem.add_memory("user-123", "assistant", "Booked the morning flight.", kind="episodic")
recent = mem.get_recent("user-123", n=10, kind="episodic")   # chronological

as_save_hook(scope) returns a save_threads_callback-compatible hook to mirror turns into episodic memory from an Agency.

Memory options

Arg	Default	Purpose
connection_string	—	MongoDB / Atlas URI (required)
database_name	agency_swarm	Database name
collection_name	memories	Collection name
vector_search_index	idx_agent_memory	MongoDB 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 old memories

Memory document shape

{
  "scope": "user-123",        // tenant / conversation / agent scope
  "kind": "semantic",         // "episodic" | "semantic"
  "role": "user",
  "content": "Prefers window seats.",
  "embedding": [ /* present only on the bring-your-own-vector path */ ],
  "ts": { "$date": "..." },
  "meta": {}
}

Demos

• demo/custom_persistence_mongo.py — Mongo-backed mirror of Agency Swarm's custom_persistence.py: run a turn, simulate a restart, verify recall.
• demo/agent_demo.py — a Gemini agent whose threads persist to Atlas, plus an MongoDB Vector Search staffing tool over a team directory (Voyage 3.5 embeddings).
• demo/memory_demo.py — MongoMemoryStore semantic + episodic recall on Atlas, runnable in both modes: bring-your-own Voyage vectors (default) or MEMORY_MODE=auto for Atlas Automated Embedding.

pip install -e ".[demo]"
pip install "openai-agents[litellm]" "litellm[proxy]"
# demo/.env: ATLAS_URI, VOYAGE_API_KEY, GEMINI_API_KEY
python demo/agent_demo.py
python demo/memory_demo.py                 # bring-your-own vectors
MEMORY_MODE=auto python demo/memory_demo.py  # Atlas Automated Embedding

Conventions

• Connection appName: devrel-integ-agencyswarm-python (server-side attribution).
• Driver handshake metadata: agency-swarm-mongodb (distinct from appName).

Tests

pip install -e ".[dev]"
pytest -q          # 25 tests, mongomock — no infra required

License

MIT