strands-mongodb-store

A MongoDB Vector Search MemoryStore for Strands Agents — semantic long-term agent memory using Automated Embedding: MongoDB generates the embeddings itself (via Voyage AI), so you store plain text and query with plain text — no external embedder, no vectors to manage.

pip install strands-mongodb-store

from strands import Agent
from strands.memory import MemoryManager
from strands_mongodb_store import MongoDBMemoryStore

store = MongoDBMemoryStore(
    name="user-memories",
    connection_string="mongodb://localhost:27017",   # self-managed, or an Atlas SRV URI
    database_name="agent", collection_name="memory",
    model="voyage-4-lite",
)
agent = Agent(memory_manager=MemoryManager(stores=[store]))

await store.add("The user prefers dark mode", metadata={"kind": "pref"})
hits = await store.search("what theme does the user like?")

How it works

• Automated Embedding. The vector index is created with a type: "autoEmbed" field and a Voyage model; MongoDB embeds your content at index-time and your query text at query-time. add stores just {_id, content, metadata, createdAt} — no vectors in your documents.
• Semantic recall via $vectorSearch ("query": <text>), ranked by vectorSearchScore, surfaced as _score.
• The vector index is created automatically if absent.

Requirements

MongoDB Vector Search with Automated Embedding, on either:

• MongoDB Atlas, or
• self-managed MongoDB Community 8.2+ running the mongot binary (Linux; Docker / tarball / package / K8s).

Automated Embedding needs a Voyage AI API key configured on the deployment (Atlas, or mongot for Community). Models: voyage-4-lite (default), voyage-4, voyage-4-large, voyage-code-3.

The memory store (strands-mongodb-store) is distinct from the byte storage backend (strands-mongodb-storage). Also published as strands-store-mongodb. 0.2.0 switched from manual embeddings to Automated Embedding (breaking).

License

MIT