Archived project — the hosted API is no longer running. Source code and full docs remain available.
Blog · LangChain

LangChain Persistent Memory API:
A Deep Dive into Cross-Session State

8 min read · LangChain · Python · Memory

LangChain is the powerhouse of the AI agent world, but its default memory components — like ConversationBufferMemory — are inherently ephemeral. They live in your server's RAM. If your script ends or your container restarts, your agent's "experience" is wiped clean.

To build production-grade applications, you need a LangChain Persistent Memory API that persists data at the database level without the overhead of managing a vector store yourself.

The Problem: The "Stateless" Agent

By default, LangChain agents treat every new session as a blank slate. While you can use SQLChatMessageHistory, it only provides a chronological log — it doesn't provide semantic context.

If a user says "Remember my API key is 12345" in session one, a standard SQL history won't help the agent "find" that fact three weeks later unless you feed the entire history back into the context window — a recipe for high costs and latency.

The Solution: Memstore as a Remote Vector Memory

By using Memstore, you can give LangChain agents a "Long-Term Memory" tool. Instead of local storage, the agent uses a REST API to store and retrieve facts.

Implementation: Custom Memstore Tool

Here is how to integrate Memstore into a LangChain agent using Python.

LangChain + Memstore integration Python 
import requests
from langchain.agents import Tool, AgentExecutor, create_openai_functions_agent
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

MEMSTORE_API_KEY = "your_memstore_key"

def remember_fact(text: str):
    url = "https://memstore.dev/v1/memory/remember"
    headers = {"Authorization": f"Bearer {MEMSTORE_API_KEY}"}
    payload = {"content": text, "metadata": {"source": "langchain_session"}}
    requests.post(url, json=payload, headers=headers)
    return "Fact stored in long-term memory."

def recall_memory(query: str):
    url = f"https://memstore.dev/v1/memory/recall?q={query}"
    headers = {"Authorization": f"Bearer {MEMSTORE_API_KEY}"}
    response = requests.get(url, headers=headers)
    memories = response.json().get("memories", [])
    if not memories:
        return "No relevant past memories found."
    return "\n".join([m['content'] for m in memories])

# Define Tools
tools = [
    Tool(
        name="store_memory",
        func=remember_fact,
        description="Use this to save important facts for future sessions."
    ),
    Tool(
        name="search_memory",
        func=recall_memory,
        description="Use this to look up past interactions or user preferences."
    )
]

llm = ChatOpenAI(model="gpt-4o")
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant with long-term memory."),
    ("placeholder", "{agent_scratchpad}"),
    ("human", "{input}"),
])

agent = create_openai_functions_agent(llm, tools, prompt)
executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

# Usage
executor.invoke({"input": "My favorite color is midnight blue. Please remember that."})

Why This Architecture Wins in 2026

In 2026, context windows are larger, but "Noise-to-Signal" ratios are still a problem. By using GET /v1/memory/recall, you only inject the relevant snippets into the prompt. This keeps your agent focused, fast, and remarkably cheap to run.

The key insight: Semantic retrieval means you only pull the 2–3 facts that matter right now — not the entire conversation history. Smaller prompts, lower costs, sharper reasoning.

This project is archived

The hosted API is no longer running. The full source — REST API, SDKs, MCP server, and Postgres schema — remains available to read.

View source on GitHub →