TruthLedger

Internet Memory: An AI-Powered Evidence Engine for Real-World Accountability

🚀 What It Does

Internet Memory is not a search engine.

It is an autonomous research agent network that continuously monitors the internet, extracts verifiable real-world outcome claims, validates them across modalities (text + video + voice), and stores them in a living accountability graph. Instead of summarizing what companies say, we track what they actually deliver and prove it.

Instead of simply showing articles that contain words like “claim” or “promise,” our system extracts real, verifiable status-based claims such as:

What a company has launched What it is actively working on What was delayed What failed What was discontinued

Every claim is backed by direct source evidence including links, snippets, and timestamps (including YouTube transcript timestamps when applicable).

The result is a structured accountability layer built on top of the internet.

🧠 The Problem

The internet forgets context.

Companies announce initiatives, launch products, delay features, or miss targets but this information gets buried in fragmented articles and videos.

There is no structured system that tracks:

Who delivered what Who delayed what What is still in progress What failed

Most systems rely on keyword matching or generic summaries, which leads to noise and misinformation.

We built a system that extracts real, evidence-backed propositions and structures them into a living knowledge graph.

🏗 How We Built It

Our system follows a multi-stage pipeline: User searches for an entity (e.g., Apple, Tesla) The system resolves ambiguity if needed Tavily retrieves high-quality web sources using expanded queries Yutori orchestrates multi-step research workflows OpenAI models extract structured status claims from full content Claims are validated and scored Evidence (URL + snippet + timestamp) is attached A dynamic knowledge graph connects entities → claims → sources YouTube videos are analyzed with transcript extraction + voice signal analysis Results update in real time

We do not generate claims without evidence.

🔎 Tavily

Multi-query web retrieval High-quality source ingestion Recall expansion when evidence is limited Structured content retrieval for downstream processing

🤖 OpenAI

Structured claim extraction Status classification (implemented / in progress / delayed / failed) Evidence validation Confidence scoring Transcript-based claim extraction from videos

🧭 Yutori

Autonomous research orchestration

Query expansion logic

Multi-step workflow coordination

Retrieval → extraction → validation pipeline management

🎙 Modulate

Voice signal analysis on video content

Detecting delivery patterns and confidence in spoken claims

Enhancing video-based evidence context

🗺 Neo4j

Knowledge graph storage

Linking entities to claims and supporting sources

Powering interactive graph visualization

🌐 Render (if applicable)

Service deployment

Scalable backend runtime

💡 Key Features

Evidence-backed claims only

Real-time scanning and updates

Disambiguation for ambiguous entities

Knowledge graph visualization

YouTube transcript + timestamp extraction

Strict filtering to avoid keyword noise

Claim categorization:

Implemented

In progress

Planned

Delayed

Failed

Deprecated

🖥 Demo Walkthrough

Search for a company (e.g., Apple)

Resolve ambiguity if necessary

Run a live scan

Watch sources being retrieved in real time

See structured claims extracted with evidence

Explore the interactive graph

View supporting articles

Analyze YouTube transcripts with timestamped evidence

The system dynamically builds structured memory from live web data.

⚙️ Technical Architecture

User → Query Expansion (Yutori) → Multi-query Retrieval (Tavily) → Content Processing → Structured Claim Extraction (OpenAI) → Validation + Scoring → Graph Construction (Neo4j) → Realtime Updates (AWS infrastructure) → Frontend Visualization

🧪 Challenges We Ran Into

Over-triggering on keyword matches (fixed with semantic claim validation)

Low recall due to strict extraction rules (fixed with two-pass retrieval expansion)

Hydration errors and real-time sync issues

Ensuring no hallucinated claims

Handling ambiguous entity names

Making sure evidence was always attached

🎯 What Makes This Different

Most AI systems summarize.

We structure and verify.

We don't just tell you what was said. We show what actually happened with evidence.

This is not another chatbot. It is a structured accountability engine.

🌍 Impact & Use Cases

Investor intelligence Competitive research Investigative journalism Regulatory monitoring Corporate accountability tracking Product roadmap tracking

🚀 Future Roadmap

Continuous monitoring with automated alerts Historical timeline view of claims Sentiment + outcome trend analysis Cross-company comparative dashboards Enterprise compliance monitoring integration

Expanded video + earnings call parsing

🏁 Final Statement

The internet forgets.

We built a system that remembers and backs every claim with evidence.