langgraph-store-dynamodb

Amazon DynamoDB long-term-memory store (BaseStore) for LangGraph — namespaced key/value agent memory with prefix search, Mongo-style filters, list_namespaces, sync + async, and native semantic search via DynamoDB vector search.

pip install langgraph-store-dynamodb

Usage

from langgraph_store_dynamodb import DynamoDBStore

store = DynamoDBStore(table_name="langgraph-store")          # table auto-created (PAY_PER_REQUEST)

store.put(("users", "1", "memories"), "food", {"text": "loves sushi", "kind": "pref"})
item  = store.get(("users", "1", "memories"), "food")
hits  = store.search(("users", "1"), filter={"kind": "pref"}, limit=10)
spaces = store.list_namespaces(prefix=("users",))

graph = builder.compile(store=store)                          # as a LangGraph store

Semantic search (DynamoDB native vector search)

Pass a LangGraph IndexConfig and the store embeds the configured fields on put and ranks search(query=...) by cosine similarity using DynamoDB's SearchVectors — no external vector database.

from langgraph_store_core import bedrock_titan_embeddings
from langgraph_store_dynamodb import DynamoDBStore

store = DynamoDBStore(
    table_name="langgraph-memory",
    index={"dims": 1024, "embed": bedrock_titan_embeddings(dimensions=1024), "fields": ["text"]},
)
store.put(("memories", "kamal"), "k1", {"text": "the user loves sushi", "kind": "pref"})
hits = store.search(("memories", "kamal"), query="what food does the user like?", filter={"kind": "pref"})
print(hits[0].score, hits[0].value)

embed may be any LangChain Embeddings, a list[str] -> list[list[float]] callable, or a provider string. fields defaults to ["$"] (the whole value as JSON). put(..., index=False) skips embedding for one item; put(..., index=["title"]) overrides the fields.

The table is created with a vector index (embedding, cosine, dims, PK as an inline filter). A vector index can only be declared at table creation, so use a new table name when enabling semantic search on an existing store. SearchConditionExpression only allows equality on a string search-schema attribute, so a prefix search resolves to its concrete namespaces (keys-only scan) and runs one ANN query per namespace, merged by score — an exact namespace is a single call. Value filters cannot be expressed there (only top-level search-schema attributes), so the store oversamples and applies them on the returned candidates.

Requires boto3>=1.43.78 and a region where DynamoDB vector search is available (GA 2026-08-05). Bedrock model access is needed only for the default Titan embedder.

Upgrading from 0.1.x

0.2.0 is a rewrite on langgraph-store-core. Same table (PK = namespace, SK = key, value, created_at, updated_at), same constructor (max_read_request_units / max_write_request_units still accepted). New in 0.2.0: prefix search (not just exact namespace), filter, offset, list_namespaces, GetOp/ListNamespacesOp in batch, ordered abatch, semantic search, and region_name / boto_session / endpoint_url options. Namespaces are now joined with a unit separator instead of :; items written by 0.1.x are still read and are migrated on their next put. aioboto3 is no longer required.

AWS permissions

dynamodb:DescribeTable, CreateTable, GetItem, PutItem, DeleteItem, Query, Scan, and SearchVectors (for semantic search), plus bedrock:InvokeModel for the default embedder.

License

MIT · part of the langgraph-store family · docs: https://skamalj.github.io/agentstate-reducer/