🧠 Pudding: Cognitive Accessibility Engine with Cognee Memory

"Your AI woke up in Vegas with no memory of last night. We fixed that."

Pudding gives every reading session a permanent, self-hosted, hybrid graph-vector memory powered by Cognee. The web finally remembers you.

The Problem

Reading online content is hard for millions of people, those with dyslexia, ADHD, and cognitive differences. But the deepest problem isn't any single session.

The real gap: Every tool forgets you the moment you close the tab. No tool learns which specific concepts you struggle with and proactively helps you across sessions.

8.4 million people with dyslexia, 6.4 million with ADHD, and hundreds of millions of non-native language readers hit a wall every single day and start from zero every single time.

The Solution

Pudding is a Chrome extension that uses Cognee's open-source hybrid graph-vector memory layer to build a permanent, evolving cognitive model of every user. It simplifies content on demand using Gemini AI, and then remembers which concepts you struggled with, forever.

	Every Other Tool	Pudding + Cognee
Memory	❌ Forgets on tab close	✅ Permanent via cognee.remember()
Recall	❌ No cross-session context	✅ Hybrid semantic + graph via cognee.recall()
Adaptation	❌ Static, no learning	✅ Gets smarter via cognee.improve()
Cleanup	❌ All or nothing	✅ Surgical removal via cognee.forget()
AI	❌ Cloud, privacy risk	✅ On-device Gemini Nano
Languages	❌ English only	✅ 10 languages

---

🧠 How Cognee Powers Pudding

Pudding uses all four operations of Cognee's memory lifecycle API:

# Every simplification session:
await cognee.remember(page_text, dataset_name="guest_pudding_memory")

# On every page load (predictive highlighting):
results = await cognee.recall("STRUGGLED_WITH concepts related to: " + keywords)

# Every 5 sessions (self-improvement):
await cognee.improve(dataset_name="guest_pudding_memory")  # prune, bridge, memify

# "I've mastered this topic, remove it":
await cognee.forget(dataset_name="guest_pudding_memory")

Why this matters:

• remember() builds a knowledge graph of every concept encountered, with entity extraction and relationship mapping
• recall() auto-routes between semantic vector search and deep graph traversal, no manual routing needed
• improve() enriches the graph over time: prunes stale nodes, bridges related concepts, adapts weights. This is the step most tools never do, and it's why Pudding gets smarter the more you use it
• forget() gives users control. "I've mastered quantum mechanics" removes it surgically, not a full wipe

Cognee is configured 100% locally: SQLite + LanceDB + Ladybug graph store. No cloud account needed. Qualifies for the Best Use of Open Source track.

---

Architecture

Chrome Extension (JavaScript)
│
├── Gemini AI (on-device)          ← simplifies text, zero privacy risk
├── Cognitive Tracker              ← scroll speed, pauses, rereads
├── Complexity Analyzer            ← scores 0-100 per paragraph
├── Predictive Highlighter         ← cognee.recall() on every page load
├── Brain Map Viewer               ← vis.js graph of your Cognee knowledge graph
├── Tune Memory button             ← triggers cognee.improve() on demand
└── Reset Memory button            ← triggers cognee.forget() surgically
     │
     ▼  HTTP (localhost:8000)
FastAPI Backend (Python)
│
├── POST /api/remember             ← cognee.remember()
├── POST /api/recall               ← cognee.recall()
├── POST /api/improve              ← cognee.improve()  ← KEY DIFFERENTIATOR
├── POST /api/forget               ← cognee.forget()
├── GET  /api/graph                ← knowledge graph for brain-map viewer
└── POST /api/check-memory         ← predictive highlight lookup
     │
     ▼
Cognee (open-source, self-hosted)
│
├── LanceDB           ← local vector store (semantic similarity)
├── Ladybug (Kuzu)    ← local graph store (relationship traversal)
└── SQLite            ← local relational store (metadata, state)

---

Features

1. Predictive Simplification (Cognee-powered)

On every page load, Pudding queries Cognee with recall() for concepts you've struggled with before and highlights them automatically, before you click anything. Red wavy underline + tooltip: "Cognee: you struggled with 'entropy' before. Click to simplify."

2. Self-Improving Memory (cognee.improve())

After every 5 simplification sessions, Pudding automatically calls cognee.improve() (memify). This post-ingestion pass prunes stale nodes, bridges related concepts, and adapts weights, so recall gets smarter over time. You can also trigger it manually with the Tune Memory button.

3. Surgical Forgetting (cognee.forget())

When you've mastered a topic, the Reset Memory button calls cognee.forget() to remove it from your struggle layer. No full wipe required. Cognee supports dataset-level surgical removal.

4. Brain Map Viewer

An interactive Vis.js graph of your entire Cognee knowledge graph:

• 🔴 Red nodes = concepts you keep struggling with (score ≥ 60)
• 🟢 Green nodes = concepts you've mastered (score < 60)
• Node size scales with difficulty
• Persists across sessions and devices (stored in Cognee's local graph DB)

5. Complexity Mapping

Colour-coded difficulty scores (0-100) across the page. Click any hard section to simplify just that part and have it remember()-ed in Cognee immediately.

6. Cognitive Adaptation Engine

Tracks scroll speed, pauses, and rereads. Auto-adjusts simplification level. All behaviour data stays local via Cognee's self-hosted storage.

7. Smart Content Restructuring

Converts dense paragraphs into structured, scannable content with key points, bullet summaries, and context.

8. Focus Mode

Blurs sidebars, hides ads, spotlights the current paragraph. Keyboard navigation with arrow keys.

9. Multilingual Support

10 languages, RTL support for Arabic, instant switching. Accessibility truly global.

---

Installation

Prerequisites

Requirement	Details
Browser	Chrome Dev/Canary ≥ 128
Python	3.11+
LLM API key	OpenAI (for Cognee entity extraction)

Step 1: Backend Setup (Cognee)

cd backend

# Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

# Install dependencies (Cognee + FastAPI)
pip install -r requirements.txt

# Configure your LLM key
cp .env.example .env
# Edit .env and set LLM_API_KEY=sk-...your-openai-key...

# Start the Cognee memory server
uvicorn server:app --reload --port 8000

Step 2: Enable Gemini Nano (optional, for on-device AI)

chrome://flags/#optimization-guide-on-device-model  → Enabled BypassPerfRequirement
chrome://flags/#prompt-api-for-gemini-nano          → Enabled
chrome://components → Optimization Guide On Device Model → Check for update

Or use the Gemini API key in options. No flags needed.

Step 3: Load the Extension

chrome://extensions/ → Enable Developer mode → Load unpacked → select repo root

Step 4: Verify

1. Open any article
2. Click Pudding → Simplify Text
3. Button shows "Saving to Brain..." → "Simplified & Saved!"
4. Click Tune Memory to trigger cognee.improve() and watch the toast: "🧠 Memory enriched by Cognee!"
5. Click View My Brain Map to see the Cognee knowledge graph in the vis.js viewer
6. Revisit a page with concepts you've simplified before. Red wavy underlines appear automatically via cognee.recall()

---

Judging Criteria Alignment

Criterion	How Pudding delivers
Potential Impact	Addresses a real problem for 100M+ people: dyslexia, ADHD, language learners all benefit from a web that remembers their cognitive journey
Creativity	First accessibility extension that uses a persistent knowledge graph to proactively adapt content before the user asks
Technical Excellence	Clean separation: Chrome extension, FastAPI, Cognee. All four memory lifecycle APIs used. Fully local, zero cloud dependencies
Best Use of Cognee	All four APIs: remember(), recall(), improve(), forget(). Self-hosted with LanceDB + Ladybug. Auto-improve after 5 sessions. Surgical forget per topic
User Experience	One-click simplification, predictive highlights, brain map, tune/reset memory, intuitive at every step
Presentation	Demo guide, architecture diagrams, lifecycle diagrams, clean README

---

Quick Demo Script

1. Open https://en.wikipedia.org/wiki/Quantum_entanglement
2. Click Pudding → Simplify Text          → concepts saved via cognee.remember()
3. Click Tune Memory                       → cognee.improve() enriches graph
4. Click View My Brain Map                 → see quantum, entanglement, entropy as red nodes
5. Open any other quantum physics article  → red wavy underlines appear automatically
6. Click Reset Memory                      → cognee.forget() clears the dataset

---

Why Open Source Cognee?

Cognee is self-hosted, open-source, and ships with built-in local backends:

• LanceDB: local vector store for semantic similarity search
• Ladybug (Kuzu-compatible): local graph store for relationship traversal
• SQLite: local relational store for metadata

Zero cloud setup required. Your cognitive graph stays on your machine. Privacy-first by design.

# That's all Cognee needs to run fully locally:
pip install cognee

---

Roadmap

• 🔐 Auth: personal Cognee cloud profiles for cross-device sync
• 📱 Cross-device sync via Cognee's managed cloud
• 🎧 Voice layer with synchronised highlighting
• 📚 Study Mode: flashcards generated from your Cognee struggle graph
• 👥 Classroom Mode: teachers can query a class's aggregate struggle graph
• 🌙 Time-based adaptation (late-night simplification mode)

---

Acknowledgements

• Cognee: the hybrid graph-vector memory layer that makes all of this possible
• Chrome AI Team: for Gemini Nano (on-device AI)
• OpenDyslexic: for the accessibility font
• Vis.js: for the brain map visualisation

---

Built for

WeMakeDevs × Cognee Hackathon 2026 "The Hangover Part AI: Where's My Context?" Track: Best Use of Open Source

Making the web remember you, so you can focus on understanding it.

⭐ Star Cognee on GitHub · Report Issues