How do I implement AsyncPostgresStore?

LangSmith Product Help
,
1

Hi LangChain community, I’m a self-professed noob and I just got started with LangMem and have been reading the docs.

However I don’t know how AsyncPostgresStore actually works. I understand that I need to create an index_config and specify an embedding model, in this caseopenai:text-embedding-3-small. I’m assuming by specifying an embedding model that there is some sort of vector table in my postgres db.

I have 3 sets of questions regarding implementation of long-term memory:

  1. Is specifying an embedding model necessary for long-term memory to work? Is it recommended? What happens if I don’t use an embedding model?
  2. I see 2 tables in my DB store and store_vectors. If I don’t specify an embedding model, then only store will be used? How do both tables get used by AsyncPostgresStore?
  3. When .aput or .asearch is called, how does memory_searcher or memory_manager figure out whether to save to store or store_vector? Throughout my setup I didn’t explicit do CREATE TABLE store or store_vector and I didn’t see any documentation on setting this up so far, so I feel very uneasy with this “magic”

Any guidance would be really appreciated!

2

  1. Embedding is optional , but it will improve the search. If you dont add embedding model into “store” object then it will only search using keyword , example : (SELECT * from store WHERE value like %query% ) . See the docs : Storage (LangGraph) | LangChain Reference

  2. If you dont specific embedding it will only use “store” table . You can check the implementation by looking on the libs itself.

  3. Unfortunately i cant answer this clearly , but all of them is well documented only inside the libs on “site-packages” of your projects.

    Here’s my implementation:

    embeddings = init_embeddings(“ollama:nomic-embed-text”)
logger.debug(f"Embeddings initialized: {embeddings is not None}")

        # Create async connection pool for better performance
        self._conn_pool = AsyncConnectionPool(
            conninfo=os.getenv("POSTGRES_URL"),
            min_size=1,
            max_size=10,
            open=False,
        )
        # Open the pool explicitly (required for newer versions)
        await self._conn_pool.open()

        # Create PostgresStore instance with vector indexing for semantic search
        self._long_term_memory_store = AsyncPostgresStore(
            conn=self._conn_pool,
            index={
                "dims": 768,
                "embed": embeddings,
                "fields": ["content"] # always put this as content or $ coz langmem store as content object , if you wrong to set this up it wont saved to store_vectors,
                "distance_type": "cosine",
            },
        )

        await self._long_term_memory_store.setup()

    CMIIW guys

3 
from langgraph.store.postgres.aio import AsyncPostgresStore - I could not find this module

langgraph version is 1.04. Help is appreciated

4

@archupsg03 try this from langgraph.store.postgres import AsyncPostgresStore

5

from langgraph.store.postgres import AsyncPostgresStore
Traceback (most recent call last):
File “”, line 1, in
ModuleNotFoundError: No module named ‘langgraph.store.postgres’

@pawel-twardziak

6

langgraph-checkpoint-postgres - installing this helped

1 Like