langgraph-store-core

Shared core for building LangGraph long-term-memory stores (BaseStore). Implement a backend by subclassing KVStore and supplying four small primitives — the core handles batch/abatch, timestamps, search filters, namespace prefix/suffix matching, list_namespaces, and semantic search.

from langgraph_store_core import KVStore

class MyStore(KVStore):
    def _read(self, prefix, key): ...        # -> row | None
    def _write(self, row): ...               # row = {prefix,key,value,created_at,updated_at,embedding}
    def _remove(self, prefix, key): ...
    def _scan(self, prefix): ...             # rows whose prefix starts with `prefix`

    # optional — native vector search; the default ranks `_scan` rows by cosine in Python
    def _vector_search(self, prefix, vector, filter, limit): ...   # -> [(row, score), ...]

Semantic search

Pass LangGraph's IndexConfig to any store built on the core and it embeds the configured fields on put and answers search(namespace, query="...") ranked by cosine similarity, with SearchItem.score populated:

store = MyStore(index={"dims": 1024, "embed": embedder, "fields": ["text"]})
store.put(("memories", "u1"), "k", {"text": "the user loves sushi"})
hits = store.search(("memories", "u1"), query="what food does the user like?")

• embed: a LangChain Embeddings, a list[str] -> list[list[float]] callable, or a provider string ("openai:text-embedding-3-small"). bedrock_titan_embeddings() is included as a ready-made callable (needs boto3).
• fields: JSON paths to embed; default ["$"] (whole value). put(..., index=False) skips one item; put(..., index=["title"]) overrides the fields.
• A store without index is filter-only and ignores query, exactly as LangGraph documents.
• Backends override _vector_search for native ANN (pgvector, Cosmos VectorDistance, Firestore find_nearest, DynamoDB SearchVectors); the core re-checks namespace and filter on what comes back, so a backend may push down only part of the filter.

langgraph_store_core.testing ships FakeEmbeddings (deterministic, no network) and MemoryKVStore for tests.

Concrete backends: langgraph-store-dynamodb, langgraph-store-postgres, langgraph-store-cosmosdb, langgraph-store-firestore.

Docs: https://skamalj.github.io/agentstate-reducer/

License

MIT