strands-dynamodb-store

A DynamoDB-native-vector MemoryStore for Strands Agents — semantic long-term agent memory backed by DynamoDB's native vector search (SearchVectors). No external vector database, no separate Storage layer — the store talks to DynamoDB directly, the way BedrockKnowledgeBaseStore talks to Bedrock.

pip install strands-dynamodb-store

from strands import Agent
from strands.memory import MemoryManager
from strands_dynamodb_store import DynamoDBMemoryStore

store = DynamoDBMemoryStore(name="user-memories", table_name="agent_memory")
agent = Agent(memory_manager=MemoryManager(stores=[store]))

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

How it works

• Semantic recall via DynamoDB search_vectors (approximate nearest-neighbor over a vector index) — ranked by similarity, at DynamoDB scale.
• You bring the embeddings. DynamoDB does not generate them; by default the store embeds with Amazon Bedrock Titan Text v2 (amazon.titan-embed-text-v2:0, 1024-dim, cosine). Pass any embedder callable to use Cohere, OpenAI, a local model, etc.
• Each add stores an item {id, content, embedding (List<Number>), metadata, createdAt}; search embeds the query and runs the ANN search.
• The table is created automatically if absent — PAY_PER_REQUEST with a vector index on the embedding attribute.

Configuration

DynamoDBMemoryStore(name, table_name, *, description=None, max_search_results=None, writable=True, extraction=None, embedder=None, dimensions=1024, distance_function="COSINE", index_name="vector_index", vector_attribute="embedding", region_name=None, boto_session=None, endpoint_url=None)

Implements the Strands MemoryStore protocol: search(query, options) and add(content, metadata).

Each result's metadata carries _score, a similarity (higher is better; for COSINE it is 1 - distance), and _distance, DynamoDB's raw value (lower is better). 0.2.0 reported the raw distance as _score; fixed in 0.2.1.

Requirements

• boto3>=1.43.78 (DynamoDB vector search, GA 2026-08-05)
• A region where DynamoDB vector search is available, plus Bedrock model access for the default embedder.

This package (the memory store) is distinct from strands-storage-dynamodb (the byte storage backend). Earlier strands-dynamodb-store releases (0.1.x) were an alias of that storage package; from 0.2.0 it is a MemoryStore.

License

MIT