Skip to content

Latest commit

 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TasteGraph

AI-Native Personal Experience Memory & Recommendation Agent

"Your personal AI concierge that remembers every great experience you've ever had."


Overview

TasteGraph is an AI-native application built on the AINative Studio platform that automatically captures, organizes, understands, and recommends the real-world experiences of its users.

Unlike Yelp, Google Maps, TripAdvisor, or public review platforms, TasteGraph builds a private lifelong memory of a user's own experiences.

Over time, the system develops a semantic understanding of the user's tastes, preferences, relationships, favorite cities, favorite restaurants, favorite events, and favorite experiences.

The result is an AI agent capable of acting like a trusted friend who has perfect memory.


Product Vision

Imagine asking:

"Where should my parents eat in Austin?"

Instead of searching Google, the AI searches years of your own memories and responds:

Based on your experiences, your parents would probably love these five restaurants. They all have quiet atmospheres, excellent seafood, outdoor seating, and you've rated each at least 4 stars.

The recommendations come from your own life, not anonymous internet reviews.


Core Philosophy

TasteGraph is built around one principle:

Your own experiences are more valuable than everyone else's reviews.

Rather than collecting public ratings, TasteGraph builds a living knowledge graph of everything you've personally experienced.

Every restaurant...

Every concert...

Every vacation...

Every coffee shop...

Every park...

Every hotel...

becomes part of your personal intelligence.


Primary Goals

The platform should:

  • Automatically remember experiences
  • Require almost zero manual effort
  • Continuously learn preferences
  • Build lifelong memory
  • Generate personalized recommendations
  • Recommend only experiences the user genuinely loved
  • Become smarter over time

Technology Stack

This project should leverage existing AINative services wherever possible to minimize custom application code.

Agent Cloud

Primary orchestration layer.

Responsibilities:

  • Long-running agents
  • Background jobs
  • Event processing
  • Scheduled workflows
  • Multi-agent coordination

ZeroDB

Primary datastore.

Use ZeroDB for:

  • Structured tables
  • Vector storage
  • Semantic search
  • Metadata
  • Event storage

ZeroMemory

Persistent AI memory.

Stores:

  • User preferences
  • Long-term facts
  • Conversation memory
  • Learned taste
  • Friend preferences

Models API

Responsible for:

  • Classification
  • Summarization
  • Embeddings
  • Extraction
  • Recommendation reasoning
  • Semantic tagging

Knowledge Graph

Stores relationships between:

User

Experiences

Cities

Venues

People

Preferences

Recommendations

Ratings


AIKit

Frontend components.

Reuse wherever possible.

Examples:

  • AI Chat
  • Timeline
  • Cards
  • Voice UI
  • Maps
  • Search
  • Recommendation Lists

Lakehouse

Used for analytics.

Examples:

  • Favorite cities
  • Favorite cuisine
  • Recommendation quality
  • Preference evolution
  • Travel history
  • Long-term insights

Product Workflow

User visits location

↓

Agent Cloud detects visit

↓

Experience Agent creates experience

↓

Models API extracts metadata

↓

ZeroDB stores structured data

↓

ZeroMemory updates long-term memory

↓

Knowledge Graph updates relationships

↓

User rates experience

↓

Taste Agent updates embeddings

↓

Future recommendations improve

Major AI Agents

1. Experience Agent

Responsibilities

  • Detect experiences
  • Merge duplicate signals
  • Create Experience records

Inputs

  • Calendar
  • GPS
  • Photos
  • Gmail
  • Reservations
  • Travel confirmations

Outputs

Experience objects


2. Memory Agent

Responsibilities

Generate structured memories.

Outputs

  • Summary
  • Metadata
  • Embeddings
  • Highlights
  • Complaints

3. Rating Agent

Responsibilities

Collect lightweight feedback.

Target completion:

Under 30 seconds.

Captures:

  • Overall Rating
  • Food
  • Service
  • Atmosphere
  • Value
  • Return Intent

4. Taste Agent

Responsibilities

Learn preferences.

Examples

User likes:

  • Outdoor seating
  • Craft cocktails
  • Quiet restaurants
  • Walkable neighborhoods

User dislikes:

  • Loud music
  • Tourist traps
  • Poor service

These become weighted preferences.


5. Recommendation Agent

Answers questions like:

Where should I eat?

Best tacos I've ever had?

Weekend in Portland?

Best date night?

Kid friendly?

Business dinner?

Parents visiting?

Uses:

  • Ratings
  • Embeddings
  • Preferences
  • Context
  • Knowledge Graph

6. Friend Agent

Stores recurring profiles.

Examples

Mom

Seafood

Gardens

Quiet

Dad

Steak

Museums

Sarah

Vegan

Coffee

Children

Parks

Ice cream

Interactive museums

Recommendations become personalized.


Core Domain Objects

The application revolves around these primary entities:

User

Experience

Venue

City

Rating

Preference

Person

Recommendation

Recommendation Feedback

Knowledge Graph Edge

Embedding

Memory

Share Session

Recommendation Philosophy

By default, recommendations should never include mediocre experiences.

Default filter:

Overall Rating >= 4

Only if explicitly requested should the AI recommend lower-rated places.

Recommendation ranking should prioritize:

  1. Rating
  2. Semantic similarity
  3. User preferences
  4. Companion preferences
  5. Occasion
  6. Recency
  7. Confidence score

Experience Lifecycle

Experience detected

↓

Metadata extracted

↓

Experience summarized

↓

Stored

↓

Embedded

↓

Knowledge Graph updated

↓

User rates experience

↓

Preferences updated

↓

Recommendation engine updated

Search Examples

Natural language search should work everywhere.

Examples

Best tacos I've ever had.

Favorite coffee in Seattle.

Restaurants my wife loved.

Every five-star restaurant in New York.

Quiet dinner with outdoor seating.

Family-friendly places in Austin.

Concerts from 2025.

Best date nights.

Friend Recommendation Examples

Example:

Parents visiting Chicago.

Agent asks:

Budget?

Walking?

Kids?

Cuisine?

Then returns:

  • 5 restaurants
  • Coffee
  • Walk
  • Dessert
  • Museums

using only experiences the user rated highly.


Rating Philosophy

Simple.

Fast.

Private.

Ratings

⭐ 1

Never again.

⭐⭐ 2

Below average.

⭐⭐⭐ 3

Good.

⭐⭐⭐⭐ 4

Recommend.

⭐⭐⭐⭐⭐ 5

Exceptional.

The rating system is not intended to compete with public reviews.

It exists to strengthen future recommendations.


Knowledge Graph

Relationships include

User

VISITED

Venue

Experience

IN_CITY

City

User

LIKES

Attribute

User

DISLIKES

Attribute

Experience

WITH

Person

Recommendation

FOR

Person

Venue

TAGGED_AS

Cuisine

Venue

TAGGED_AS

Atmosphere

MVP Scope

Version 1 should include:

✅ Authentication

✅ User onboarding

✅ Experience detection

✅ AI summaries

✅ Timeline

✅ Ratings

✅ Recommendation engine

✅ Taste learning

✅ Friend profiles

✅ City pages

✅ AI chat

✅ Share recommendations

Everything else should be deferred until after MVP.


Future Roadmap

  • Apple Watch integration
  • Oura integration
  • WHOOP integration
  • Camera vision
  • Receipt OCR
  • Menu recognition
  • Trip planner
  • AI travel guide
  • Collaborative recommendations
  • Taste similarity
  • Public recommendation publishing
  • Voice-first interaction
  • Offline mode

Engineering Principles

  • Build on existing AINative APIs before writing custom code.
  • Treat every feature as an Agent Cloud workflow where possible.
  • Use ZeroDB as the single source of truth for structured and vector data.
  • Store durable user preferences in ZeroMemory and expose them through semantic retrieval.
  • Model relationships in the Knowledge Graph instead of duplicating data.
  • Favor event-driven workflows over synchronous processing.
  • Design agents to be idempotent and composable.
  • Keep AI prompts versioned and configurable.
  • Every experience should improve future recommendations.

Definition of Done

A feature is considered complete when:

  • It is orchestrated through Agent Cloud where appropriate.
  • Data is persisted in ZeroDB using the agreed schema.
  • Relevant memories are written to ZeroMemory.
  • Knowledge Graph relationships are updated.
  • Embeddings are generated for semantic search.
  • AI reasoning uses structured data before free-text prompts.
  • The frontend uses AIKit components whenever available.
  • Events are emitted for analytics and Lakehouse ingestion.
  • Tests cover business logic and agent workflows.
  • Documentation is updated.

Mission Statement

TasteGraph transforms a lifetime of experiences into a living, continuously learning AI memory. Every meal, trip, concert, hike, hotel, and conversation enriches a personal knowledge graph that becomes more valuable with time. By combining explicit ratings, semantic memory, vector search, and autonomous agents, TasteGraph evolves into a trusted personal concierge capable of delivering recommendations that feel less like search results and more like advice from your future self.

About

AI-Native Personal Experience Memory & Recommendation Agent

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors