​
LangChain example

from langchain.memory import ConversationBufferMemory
from langchain.chat_models import ChatOpenAI
import os, requests

MEMORY_API = "https://api.flumes.ai/v1"
HEADERS = {"Authorization": f"Bearer {os.getenv('FLUMES_API_KEY')}", "Content-Type": "application/json"}

class FlumesChatMemory(ConversationBufferMemory):
    """Custom Memory class that persists to Flumes after every step."""

    def save_context(self, inputs, outputs):
        super().save_context(inputs, outputs)
        requests.post(
            f"{MEMORY_API}/memories",
            headers=HEADERS,
            json={
                "user_id": inputs.get("session_id"),
                "messages": [
                    {"role": "user", "content": inputs["input"]},
                    {"role": "assistant", "content": outputs["response"]},
                ],
            },
        )

    def load_memory_variables(self, inputs):
        r = requests.get(
            f"{MEMORY_API}/memories",
            headers=HEADERS,
            params={"user_id": inputs.get("session_id"), "limit": 10},
        )
        memories = "\n".join([m["memory"] for m in r.json()])
        return {self.memory_key: memories}

llm = ChatOpenAI(model="gpt-4o", temperature=0)
agent_memory = FlumesChatMemory(memory_key="history")
Use the limit param to prevent blowing past the model context window.

​
Manual prompt injection

If you don’t use a framework, just hydrate the chat history manually: 
const past = await fetch("https://api.flumes.ai/v1/memories?user_id=session-123&limit=10", {
  headers: { Authorization: `Bearer ${process.env.FLUMES_API_KEY}` }
}).then(r => r.json())

const systemPrompt = `You are a helpful assistant. Here is what the user told you previously:\n${past.map(m => m.memory).join("\n")}`