DuplicatePreparedStatement error when chaining agents with PostgresSaver

[wlf32]

Hi everyone,

I’m building a three-stage LangGraph bot—intake_agent → planner_agent → research_agent—and using Supabase Postgres to persist memory via PostgresSaver. My single-agent prototype worked perfectly, but during the second hand-off (planner → research) I now see:

psycopg.errors.DuplicatePreparedStatement: prepared statement "_pg3_0" already exists

I’ve tried connection pools, disabling auto-prepare, and even unique_name=True, but nothing prevents the clash. Occasionally this is followed by a BadRequestError from OpenAI about missing ToolMessage responses, despite wiring the ToolNode and edges as in the tutorials.

Here is my current environment:
• langgraph-checkpoint-postgres 2.0.21
• psycopg 3.2.7
• langchain-openai 0.3.5
• Python 3.13 on macOS Sequoia

And here is my get_memory helper—note that I added these extra options for debugging; my initial version was much simpler:

from psycopg_pool import ConnectionPool
from langgraph.checkpoint.postgres import PostgresSaver
import os

def get_memory():
    pool = ConnectionPool(
        conninfo=os.environ["DB_URI"],
        open=True, # warm up one connection
        max_size=10, # allow up to 10 concurrent sessions
        kwargs={
            "autocommit": True, # commit each checkpoint immediately
            "prepare_threshold": None # disable auto-PREPARE to avoid name collisions
        }
    )
    return PostgresSaver(pool) # positionally pass pool for v2.0.x

Despite these tweaks, the second hand-off still crashes. Has anyone overcome this in a multi-agent LangGraph setup? Would you recommend a different persistence store—Redis, S3, Qdrant—or a different graph architecture to avoid the pipeline/prepared-statement clash? Any suggestions or pointers would be hugely appreciated.

Thanks in advance!

Nick

[lukajose]

I am facing this issue, did you find a solution ? I am trying to use pgbouncer with transactions. I find that if you use postgres connections directly or session in pgboucer it works. It doesnt scale well as you have a connection bottleneck.