Inbrain

The first Self-Evolving Prediction Intelligence Network.

Inbrain is an AI-driven social prediction infrastructure built on top of GBrain's knowledge graph and long-term memory system. It transforms raw social signals into high-quality prediction markets with autonomous reasoning, continuous learning, and multi-agent collaboration.

Traditional prediction markets are amnesiac — no memory, no context, no learning. Inbrain gives prediction markets a brain.

What makes Inbrain different

Architecture

Signal Agent → Brain Agent → Execution Agent → Analyst Agent
     ↑                                              ↓
     └──────────── Dream Cycle (nightly) ←──────────┘

Four-Layer Agent System

Signal Agent — Real-time event detection

• Monitors X (Twitter), news, on-chain data, Polymarket, Kalshi
• Filters by engagement velocity and KOL influence
• Emits structured prediction signals

Brain Agent — Central reasoning engine

• Queries Inbrain Memory Graph for historical context
• AI quality evaluation (verifiability, historical similarity, community potential)
• Cross-platform crowd wisdom integration
• Only markets scoring above threshold get created

Execution Agent — Market lifecycle management

• Continuous monitoring of active markets
• Real-time probability updates based on new evidence
• Auto-resolution when verifiable outcomes are found
• Brier score tracking for calibration

Analyst Agent — Intelligence output

• Daily AI reports with top markets and alpha opportunities
• Weekly trend summaries
• Alpha discovery (AI vs. crowd divergence)
• Distribution to Discord, X, and community

Dream Cycle (Nightly Self-Evolution)

Every night, Inbrain runs a five-phase self-improvement cycle:

1. Full Review — Analyze all markets and prediction accuracy
2. Forecast Generation — Predict tomorrow's high-potential markets
3. Knowledge Gap Discovery — Find and fill missing data
4. Meta-Model Update — Learn calibration patterns (e.g., "KOL tweets in bull markets have 23% higher accuracy")
5. Community Publishing — Auto-distribute intelligence reports

Quick Start

# Clone and install
git clone https://github.com/inbrain-ai/inbrain.git
cd inbrain
bun install

# Initialize local brain (2 seconds, no Docker)
bun run dev init --pglite

# Start the prediction engine
bun run dev start

# Run a one-off dream cycle
bun run dev dream

Environment Variables

# Required: AI provider
OPENAI_API_KEY=sk-...
# or
ANTHROPIC_API_KEY=sk-ant-...

# Optional: Signal sources
INBRAIN_X_BEARER_TOKEN=...        # Twitter/X API
INBRAIN_NEWS_API_KEY=...          # NewsAPI

# Optional: Distribution
INBRAIN_DISCORD_WEBHOOK=...       # Discord reports

Core Capabilities

From GBrain Foundation

• Hybrid search — Vector + BM25 + knowledge graph traversal
• Self-wiring knowledge graph — Auto-extracted entity relationships
• 43 curated skills — Signal capture, enrichment, querying, brain ops
• PGLite or Postgres — Zero-config local or production-scale

Inbrain Prediction Layer

• AI Market Quality Engine — Filters out low-quality, unverifiable markets
• Cross-Platform Crowd Wisdom — Polymarket + Kalshi + PredictIt consensus
• Auto-Resolution — AI-verified market settlement with evidence
• Prediction Meta-Model — Self-improving calibration from historical outcomes
• Alpha Discovery — Finds markets where AI significantly diverges from crowd

How it works

Real-time Signal → AI Evaluation → Market Creation → Dynamic Monitoring
       → Auto-Resolution → Learning Reinforcement → Better Predictions

1. Signal Agent detects a trending tweet about ETH ETF approval
2. Brain Agent queries memory: "similar events historically resolved at 61% YES"
3. Brain Agent scores quality (verifiability: 85, historical similarity: 72, community: 90)
4. Market is created: "Will ETH ETF be approved by Q1 2027?" with AI estimate 58% YES
5. Execution Agent monitors news, updates odds as SEC statements emerge
6. When outcome is confirmed, auto-resolution triggers with evidence links
7. Dream Cycle records: "crypto regulation predictions are overconfident by 8%"
8. Next similar market is automatically better calibrated

Tech Stack

• Runtime: Bun
• Language: TypeScript
• Database: PGLite (local) / PostgreSQL + pgvector (production)
• AI: Claude, GPT-4, Gemini via AI SDK
• Knowledge Graph: Auto-wiring entity extraction
• Protocol: MCP (Model Context Protocol)
• Data Sources: Twitter API, NewsAPI, Polymarket API, Kalshi API

Contributing

bun run test          # fast unit tests
bun run verify        # pre-push checks
bun run typecheck     # TypeScript validation

License

MIT — Built on GBrain by Garry Tan (YC President).

Vision

Inbrain aims to be the global AI-native Prediction Intelligence Network — not just for crypto prediction markets, but expandable to:

• Political prediction
• Financial event forecasting
• AI agent markets
• Social trend prediction
• Real-time information verification
• Decentralized Oracle Intelligence

The goal: upgrade prediction markets from betting tools into next-generation social intelligence infrastructure.