Production agents with persistent memory

Because production agents need production memory

The Problem

The OpenAI Agents SDK is the real deal. Production-ready. Battle-tested. The official way to build agents with OpenAI models.

It handles tool calls, retries, streaming, and complex workflows. Your agents work reliably in production.

But they still wake up every morning not knowing who they are.

Your production agent handles thousands of requests flawlessly. Each one in complete isolation. No memory of what it learned yesterday. No recall of the patterns it discovered. Every request is day one.

It's like having a customer service team that's incredibly competent—but each agent has never heard of your company before.

How Stompy Helps

Stompy gives your OpenAI agents the memory they deserve.

With built-in MCP support via SSE transport, your agents gain:

- Persistent context: Project knowledge survives between sessions - Learned patterns: What worked yesterday informs today - User preferences: Remember what users have told you before - Domain expertise: Accumulate knowledge over time

The SDK handles production reliability. Stompy handles production memory. Your agents finally remember.

Integration Walkthrough

Import MCPServerSse from the SDK

The OpenAI Agents SDK has built-in MCP support with SSE transport.

from agents import Agent, Runner
from agents.mcp import MCPServerSse
import os

Connect Stompy via SSE with bearer auth

Use async context manager to connect to Stompy and get memory tools automatically.

async with MCPServerSse(
    name="Stompy Memory",
    params={
        "url": "https://mcp.stompy.ai/sse",
        "headers": {"Authorization": f"Bearer {os.environ['STOMPY_TOKEN']}"}
    },
) as stompy:
    # Create agent with persistent memory
    agent = Agent(
        name="Assistant",
        instructions="You have persistent memory via Stompy. Use it.",
        mcp_servers=[stompy],
    )

Agent saves knowledge with lock_context

Your agent learns something important? It saves it. Permanently.

# Agent helping with a project
result = await Runner.run(
    agent,
    "Set up the database schema for our e-commerce app"
)

# Agent saves the schema decision:
# → lock_context(topic="database_schema",
#     content="PostgreSQL with: users, products, orders, order_items.
#              UUID primary keys. Created/updated timestamps on all.
#              Soft deletes via deleted_at. Full-text search on products.",
#     priority="always_check",
#     tags="database,schema,postgresql"
# )
# → Creates v1.0

Agent recalls in every future session

Days or months later, your agent knows exactly where you left off.

# Session 30: Agent remembers everything
result = await Runner.run(
    agent,
    "Add a reviews table to the database"
)

# Agent calls: recall_context("database_schema") → v1.0
# Agent: "Adding reviews table to your PostgreSQL schema.
#   Following your conventions: UUID primary key, created_at/
#   updated_at timestamps, soft delete via deleted_at.
#   Foreign keys to users and products tables.
#
#   Here's the migration..."

# Production memory. Zero re-explaining.