Your SDLC shouldn't have amnesia.

# Clone the repo
git clone https://github.com/your-org/AMPM.git
cd AMPM

# Create virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Install Playwright browsers (for Google Meet)
playwright install chromium

# macOS: Install audio dependencies
brew install portaudio ffmpeg

cp .env.example .env

Edit .env with your API keys:

• OPENAI_API_KEY - platform.openai.com (for Whisper STT)
• CEREBRAS_API_KEY - cloud.cerebras.ai (for fast LLM)
• ELEVENLABS_API_KEY - elevenlabs.io (for TTS)

Web App (recommended for first-time use):

./run_webapp.sh

Access at http://localhost:8501

Local Voice Bot (microphone + speakers):

./run_voice.sh

Google Meet Bot (join a meeting):

./run_meeting.sh "https://meet.google.com/abc-defg-hij"

Note: All scripts automatically activate the virtual environment. See SCRIPTS.md for details.

AMPM is a contextual memory system for software development workflows. It connects meetings, PRDs, Blueprints, Work Orders, and Code in a knowledge graph—so teams can instantly understand WHY decisions were made and see the ripple effects of changes in real-time.

Two core capabilities:

1. Context Retrieval - Ask "Why did we decide X?" and get an instant answer with sources
2. Ripple Detection - Change a PRD and instantly see all downstream artifacts affected

As work moves from PRD → Blueprint → Work Order → Code, context dies.

• Engineers waste 5-10 hours/week searching for "why"
• Teams build the wrong thing because intent got lost
• Changes upstream don't ripple downstream—until it's too late

./run_webapp.sh

The web interface provides:

• Ask Questions: Natural language queries with voice input option
• Add Info: Real-time updates during meetings
• Decision Ledger: Browse all decisions with search
• Action Items: Track tasks by person and status
• Meeting History: Full meeting timeline
• Blockers: Monitor impediments

Access at http://localhost:8501

./run_voice.sh

Project: ShopFlow E-commerce Platform Feature: User Authentication

Ask AMPM:

• "Why are we using OAuth instead of SAML?"
• "What happens if we change the auth requirement?"
• "Who decided on Passport.js?"

Response time: ~2 seconds

Demo Mode - Type questions in terminal:

./run_meeting.sh "https://meet.google.com/abc-defg-hij"

Live Audio Mode - Listens to meeting audio:

./run_meeting.sh "https://meet.google.com/abc-defg-hij" --live

AMPM/
├── ampm/                    # Main package
│   ├── core/               # Core functionality
│   │   ├── graph.py        # Meeting knowledge graph
│   │   ├── embeddings.py   # Vector search
│   │   └── query.py        # Natural language queries
│   ├── agents/             # Agent behaviors
│   ├── interfaces/         # User interfaces
│   │   ├── voice_bot.py    # Local voice bot
│   │   └── meet_bot.py     # Google Meet bot
│   └── ingest/             # Data loading
├── data/
│   └── samples/            # Sample meeting data
├── docs/                   # Documentation
│   ├── architecture/
│   └── pitch/
├── tests/
├── app.py                  # Streamlit web app
├── ampm_logo.png           # AMPM logo
├── run.py                  # Run local voice bot
├── run_meet.py             # Run Google Meet bot
├── run_webapp.sh           # Launch web app
├── run_voice.sh            # Launch voice bot
├── run_meeting.sh          # Launch meeting bot
├── test_webapp.sh          # Test webapp setup
├── requirements.txt
├── .env.example
└── README.md

AMPM comes with sample meeting data from a fictional project:

Project: ShopFlow E-commerce Platform Team: Sarah (PM), Mike (CTO), Bob (Eng Lead), Alice (Design) Timeline: May 1-9, 2026

Key Events:

• Sprint Planning: Decided to use Stripe over Braintree
• Legal approval blocker (took 8 days instead of 3)
• Leadership escalation resolved the blocker

Try asking:

• "Why did we choose Stripe?"
• "What happened with Legal?"
• "Who approved the payment integration?"

from ampm import AMPMBot

bot = AMPMBot()

# Ask a question (text only)
response = bot.ask("Why did we choose Stripe?")
print(response)

# Ask and speak the response
response = bot.ask_and_speak("Who made the checkout decision?")

# Run interactive loop
bot.run()

from ampm import MeetingKnowledge

# Load custom meeting data
knowledge = MeetingKnowledge("path/to/meetings.json")

# Get context for LLM
context = knowledge.get_context()
print(f"Project: {knowledge.project_name}")
print(f"Meetings: {knowledge.meeting_count}")

import asyncio
from ampm import DemoMeetBot, MeetBot

# Demo mode (type questions)
bot = DemoMeetBot("https://meet.google.com/xxx-xxxx-xxx")
asyncio.run(bot.start())

# Live audio mode
bot = MeetBot("https://meet.google.com/xxx-xxxx-xxx")
asyncio.run(bot.start())

• Check your microphone is connected
• On macOS, grant terminal microphone permission in System Preferences

• Verify your CEREBRAS_API_KEY is correct
• Check your internet connection

• Verify your ELEVENLABS_API_KEY is correct
• Check you have remaining character quota

• Use headphones instead of speakers
• The bot pauses listening while speaking

The bot needs virtual audio routing to speak in Google Meet.

The bot plays audio through your system, but Google Meet needs to receive it through a microphone input. You need to route the bot's audio output to a virtual microphone that Google Meet can use.

macOS Setup (Recommended):

1. Install BlackHole: brew install blackhole-2ch or download from GitHub
2. Open Audio MIDI Setup (Applications → Utilities)
3. Click + → Create Multi-Output Device
4. Check both:
   • ✅ Built-in Output (so you can still hear)
   • ✅ BlackHole 2ch (so Google Meet can capture)
5. Set this Multi-Output Device as your System Output (System Preferences → Sound → Output)
6. In Google Meet, go to Settings → Audio
7. Set Microphone to BlackHole 2ch
8. Now when the bot speaks, audio goes: Bot → BlackHole → Google Meet mic ✅

Windows Setup:

1. Install VB-Audio Virtual Cable
2. Set it as your default playback device
3. In Google Meet settings, set microphone to "VB-Audio Virtual Cable"

Linux Setup:

1. Install PulseAudio: sudo apt install pulseaudio
2. Create loopback: pactl load-module module-loopback
3. Configure Google Meet to use loopback as microphone

• You need to be invited to the meeting
• Log into the correct Google account when prompted
• Consider using a separate Google account for the bot

Product teams have meetings all day. Critical decisions happen. Action items are assigned.

Then it all disappears.

• 23 hours/month per person in meetings (Atlassian 2024)
• 40% of that time wasted due to lack of context (HBR)
• Critical knowledge walks out the door when PMs leave

AMPM builds a living memory of your product decisions:

• Decision Memory: "Why did we decide that?" → Instant answer with full context
• Cross-Meeting Tracking: Follow action items across all meetings
• New PM Onboarding: Day 1 productivity instead of week 3
• Meeting ROI: Data-driven meeting optimization

MIT

Your SDLC shouldn't have amnesia. We fixed that.