Memory Agent Demo

A small conference demo for "Building AI Agents That Remember." It compares the same Acme Dashboard support question in two modes:

• With Memory: MongoDB + Pinecone + OpenAI memory pipeline
• Without Memory: a stateless baseline that answers from the current message only

The browser UI sends the same prompt to both modes in parallel so the contrast is easy to narrate live.

Prerequisites

• Node.js 18+
• A MongoDB instance
• A Pinecone index
• OpenAI API access

Setup

1. Install dependencies:

npm install

2. Copy .env.example to .env and fill in your keys:

cp .env.example .env

3. Create a Pinecone index that matches your embedding model dimensions. For text-embedding-3-small, use 1536 dimensions.

Run The Demo

Seed Sarah's demo history:

node seed.js

Start the server:

node server.js

Open the app:

http://localhost:3000

On-Stage Flow

Try these prompts with sarah_demo:

• Hey, any news on my CSV issue?
• By the way, we just upgraded to the Enterprise plan.
• Can you remind me what the export row limit was?

Then switch the user to new_user_123 and ask:

• Hi, I need help with my account.

Backend-Only Curl Examples

Memory-enabled:

curl -X POST http://localhost:3000/chat \
  -H "Content-Type: application/json" \
  -d '{"userId":"sarah_demo","message":"Hey, any news on my CSV issue?","mode":"memory"}'

Stateless baseline:

curl -X POST http://localhost:3000/chat \
  -H "Content-Type: application/json" \
  -d '{"userId":"sarah_demo","message":"Hey, any news on my CSV issue?","mode":"stateless"}'

Memory Pipeline

1. Capture: save messages to MongoDB and Pinecone
2. Retrieve: find relevant memories and load the user profile
3. Summarize: turn recent conversations into structured memory
4. Evolve: decay old memories and detect contradictions

Demo Notes

• The response includes memoryUsed so you can show the retrieved context.
• Important pipeline steps log to the terminal with labels like [CAPTURE], [RETRIEVE], and [EVOLVE].
• This is intentionally a demo app, not a production system.