Conversations that continue across sessions

Because agent conversations shouldn't have amnesia

The Problem

AutoGen agents converse brilliantly. They debate, iterate, and reach conclusions through dialogue.

And then: session ends, conversation forgotten.

Your agents had a breakthrough insight yesterday. Today they're starting from "Hello, I'm an AI assistant." The conversation that took 20 rounds of refinement? Gone. The consensus they reached? What consensus?

It's like having a productive meeting with no notes and no memory of what was discussed.

How Stompy Helps

Stompy preserves the outcomes of agent conversations, so future sessions build on past conclusions.

Your AutoGen setup gains: - Conversation outcomes persist: Decisions and conclusions carry forward - Agent specializations deepen: Each agent remembers their developed expertise - Debate history informs: Agents know what's already been discussed - Consensus endures: Hard-won agreements don't need to be re-negotiated

Integration Walkthrough

Install autogen-ext with MCP support

AutoGen has official MCP integration built into autogen-ext. Just 2 lines to add any MCP server.

pip install -U "autogen-ext[mcp]"

Add Stompy to your agent via SSE

Use SseServerParams with bearer auth to connect to Stompy. mcp_server_tools() gets all memory tools.

from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.tools.mcp import SseServerParams, mcp_server_tools
import os

# Connect to Stompy via SSE (2 lines!)
stompy_server = SseServerParams(
    url="https://mcp.stompy.ai/sse",
    headers={"Authorization": f"Bearer {os.environ['STOMPY_TOKEN']}"}
)
tools = await mcp_server_tools(stompy_server)

# Create agent with persistent memory
agent = AssistantAgent(
    name="architect",
    model_client=OpenAIChatCompletionClient(model="gpt-4o"),
    tools=tools,
    reflect_on_tool_use=True
)

Agents save debate conclusions with lock_context

After productive multi-agent debate, save the consensus. Architecture decisions persist forever.

# Session 1: Agents debate system design
result = await agent.run(
    task="Design the order system architecture"
)
# After 20 rounds of productive debate...

# Agent calls lock_context:
# lock_context(topic="order_system_design",
#   content="Architecture: Event sourcing + CQRS.
#            Order aggregate emits OrderCreated, ItemAdded,
#            OrderSubmitted events. PostgreSQL event store
#            with outbox pattern for reliability.",
#   priority="always_check", tags="architecture,orders")
# → Creates v1.0

Future sessions build on past decisions

Days or months later, agents recall previous debates. No re-litigation of settled decisions.

# Session 2 (or month later): Decisions persist
result = await agent.run(
    task="Implement the order aggregate"
)

# Agent calls: context_search("order architecture")
# Agent: "Implementing Order aggregate with event sourcing
#   as we designed. Creating OrderCreated, ItemAdded events
#   with the outbox pattern we agreed on.
#
#   Here's the implementation..."

# No more "What architecture pattern for orders?"
# The debate is settled. Now we build.