Edamame

AI clones for organizational memory. Edamame ingests knowledge from Slack, Google Drive, Gmail, GitHub, Notion, and Jira to create digital twin clones of every person in your organization - queryable 24/7 in text or voice, with source citations and continual learning.

Built at TreeHacks 2026.

Features

• Clone Chat — Talk to any employee's digital twin in text or voice. Personality-aware responses with inline source citations. Clones learn from every conversation via fact extraction and episodic memory.
• CEO Insights — Multi-clone sentiment analysis. Ask a strategic question and poll all clones simultaneously for per-person stances, confidence scores, and aggregated themes. Animated agent network visualization shows the query in real-time.
• Onboarding Briefs — Auto-generated "here's what you need to know" docs for new hires: key people, recent decisions, open risks.
• Offboarding Handoff Packs — When someone leaves, their clone generates ownership areas, unresolved work, key links, and suggested new owners.
• Knowledge Base — Semantic search across all ingested memories with source and type filtering.
• Clone-to-Clone Consultation — When a clone doesn't know something, it consults other clones via an agent-to-agent protocol.
• Voice I/O — Full spoken conversations with clones using Whisper (STT) and OpenAI TTS.
• Real-time Slack Learning — Webhook-driven ingestion. Clones absorb new Slack messages as they're sent.
• Synthetic Data Generation — Seeded deterministic generator creates realistic Slack messages, Drive docs, GitHub commits, emails, Jira tickets, and Notion pages for demos.

Architecture

├── frontend/              Next.js 16 app (UI + API routes)
│   ├── app/               Pages and API routes
│   │   ├── page.tsx       Landing / auth page
│   │   ├── ceo/           CEO view (insights, clones, knowledge)
│   │   ├── employee/      Employee view (chat, coworkers, knowledge)
│   │   ├── (app)/         Dashboard, settings, clone management
│   │   └── api/
│   │       ├── edamame/  Clone chat, insights, onboarding, offboarding
│   │       ├── chat/      General chat endpoint
│   │       ├── voice/     Whisper transcription + TTS synthesis
│   │       ├── ingest/    Data ingestion + synthetic generation
│   │       ├── memory/    Memory search + compaction
│   │       ├── clones/    Clone CRUD
│   │       ├── slack/     Slack OAuth, sync, webhook events
│   │       ├── google-drive/ Drive sync
│   │       ├── gmail/     Gmail sync
│   │       ├── github/    GitHub sync
│   │       ├── notion/    Notion sync
│   │       └── auth/      Google OAuth flow
│   ├── components/
│   │   ├── edamame/      InsightsView, ClonesView, EmployeeChatView,
│   │   │                  KnowledgeView, AgentNetworkView, Sidebars
│   │   ├── chat/          ChatWindow, MessageBubble, ThinkingPanel
│   │   ├── voice/         VoiceButton, Waveform
│   │   ├── dashboard/     CloneGrid, CloneCard, ConversationLog
│   │   └── clone-builder/ PersonalityForm, DocumentUpload
│   └── lib/
│       ├── agents/        OpenAI client, clone-brain prompting,
│       │                  collaboration (clone-to-clone), Perplexity
│       ├── core/          Types, Supabase client, chunker, utils
│       ├── integrations/  Slack, Google, GitHub, Notion connectors
│       ├── memory/        Frontend memory search helpers
│       └── edamame/      Edamame API client + types
├── backend/
│   ├── memory/            Memory system: retrieval, compaction,
│   │   │                  continual learning, episodic extraction
│   │   └── synthetic/     Synthetic data generators (Slack, Drive,
│   │                      email, GitHub, Jira, Notion, world builder)
│   ├── modal/             Python modules for Modal deployment
│   │                      (agent, embed, STT, TTS)
│   └── supabase/          SQL schema + migrations
└── synthetic_corpus/      Pre-generated demo data

Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS
Database	Supabase (PostgreSQL + pgvector)
LLM	OpenAI GPT-4o
Embeddings	text-embedding-3-small (1536-d), IVFFlat indexing
Voice	Whisper-1 (STT), TTS-1 with Nova voice
Integrations	Slack API, Google OAuth (Drive + Gmail), Octokit (GitHub), Notion API
ML Infra	Modal (optional, for hosted inference)
Memory	Supabase (primary), Mem0 (optional fallback)

Prerequisites

• Node.js 20+
• npm
• Supabase project with pgvector extension
• OpenAI API key
• Python 3.11+ (only for backend/modal/)

Quick Start

1. Install dependencies

cd frontend
npm install

2. Configure environment

Copy .env.example to frontend/.env.local and fill in your keys (OpenAI, Supabase, and optionally Google OAuth, Slack, GitHub, Notion).

3. Initialize database

Run backend/supabase/schema.sql in your Supabase SQL editor. This creates:

• clones — One per person, includes personality and expertise tags
• memories — Unified knowledge store with type discriminators (document, chunk, fact, snapshot, category, episodic) and source tags (slack, gdrive, email, github, notion, jira, voice, conversation)
• messages — Flat chat history grouped by conversation
• integrations — OAuth credentials and sync config
• match_memories — pgvector cosine similarity search function

If migrating from an older schema, use backend/supabase/migrate.sql instead.

4. Run the app

cd frontend
npm run dev

Open http://localhost:3000.

5. (Optional) Generate synthetic data

Hit POST /api/ingest/synthetic with a clone ID to populate a clone with realistic demo data across all sources.

How Memory Works

All clone knowledge lives in a single memories table:

1. Ingestion — Data from Slack, Drive, Gmail, GitHub, Notion, Jira is synced and chunked (500-token segments with 50-token overlap)
2. Embedding — Each chunk/fact gets a 1536-d embedding via text-embedding-3-small
3. Retrieval — Semantic vector search via match_memories RPC with keyword fallback. Results are re-ranked using a composite score: similarity + recencyBonus(occurred_at)
4. Continual Learning — Conversations trigger fact extraction and episodic memory extraction. Near-duplicates (similarity > 0.88) are reinforced instead of duplicated
5. Compaction — Weekly summarization rolls up stale facts into category summaries. Monthly rewind creates snapshots

Scripts

From frontend/:

Command	Description
npm run dev	Start dev server
npm run build	Production build
npm run start	Run production server
npm run lint	Run ESLint

Modal (Python Backend)

backend/modal/ contains optional Python modules for Modal-hosted inference:

• agent.py — Clone reasoning
• embed.py — Embedding generation
• stt.py — Speech-to-text
• tts.py — Text-to-speech
• multi_agent.py — Multi-agent orchestration

Set MODAL_BASE_URL in your env to point to the deployed Modal service.

cd backend/modal
pip install -r requirements.txt