Skip to content

Strands Memory Stores

Real Strands MemoryStore implementations — long-term, semantic agent memory (search / add) — each backed by its datastore's native vector search. No external vector database.

Why native vector search (not the Storage interface)?

A MemoryStore needs search(query) -> ranked entries, but strands.storage.Storage has no search primitive (only write/read/delete/list). So a semantic store can't ride generically on Storage — it must use a backend that has search. Each store below talks to its backend directly, the way the SDK's BedrockKnowledgeBaseStore talks to Bedrock.

The stores

Package Vector engine Backend requirement
strands-dynamodb-store DynamoDB native SearchVectors boto3>=1.43.78
strands-postgres-store · strands-store-postgres PostgreSQL pgvector (<=> / HNSW) vector extension on the server
strands-mongodb-store · strands-store-mongodb MongoDB Vector Search — Automated Embedding (Voyage) Atlas or self-managed Community 8.2+ (mongot) + a Voyage key

Usage

from strands import Agent
from strands.memory import MemoryManager
from strands_postgres_store import PostgresMemoryStore   # or DynamoDBMemoryStore / MongoDBMemoryStore

store = PostgresMemoryStore(name="user-memories", url="postgresql://user:pass@host:5432/db")
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?")
for h in hits:
    print(h.metadata["_score"], h.content)

Constructor per backend:

from strands_dynamodb_store import DynamoDBMemoryStore
DynamoDBMemoryStore(name="mem", table_name="agent_memory")   # native SearchVectors

from strands_postgres_store import PostgresMemoryStore
PostgresMemoryStore(name="mem", url="postgresql://user:pass@host:5432/db")

from strands_mongodb_store import MongoDBMemoryStore
MongoDBMemoryStore(name="mem",
    connection_string="mongodb+srv://user:pass@cluster.mongodb.net",  # or self-managed URI
    database_name="agent", collection_name="memory", model="voyage-4-lite")

How it works

  • Semantic recall via the backend's native ANN — ranked by similarity, surfaced as _score in each entry's metadata.
  • Embeddings. DynamoDB and Postgres bring their own — by default each embeds with Amazon Bedrock Titan Text v2 (1024-dim, cosine); pass any embedder callable to use Cohere / OpenAI / a local model. MongoDB uses Automated Embedding — MongoDB (via Voyage AI) generates embeddings at index- and query-time, so you store and query plain text and there is no external embedder.
  • Each add stores a record (content + metadata; DynamoDB/Postgres also store the vector); the table / index is created automatically.

Store vs. storage

A memory store (strands-<backend>-store) is the semantic layer. The byte Storage backend is strands-<backend>-storage / strands-storage-<backend> — a different, lower layer.

Backend notes

  • DynamoDB — native vector search GA'd 2026-08-05; needs a region where it's available. Table is PAY_PER_REQUEST with a vector index.
  • PostgreSQL — needs the pgvector extension (apt install postgresql-16-pgvector, brew install pgvector, or the extension on RDS / Cloud SQL / Azure). The store runs CREATE EXTENSION IF NOT EXISTS vector.
  • MongoDB — MongoDB Vector Search on Atlas or self-managed Community 8.2+ running the mongot binary (Linux; Docker / tarball / K8s). Automated Embedding needs a Voyage AI API key configured on the deployment. Models: voyage-4-lite (default), voyage-4, voyage-4-large, voyage-code-3.

License

MIT · part of the strands-agents-session family.