The Technology Behind Riera

Riera is a graph-based multi-agent LLM system that transforms natural language into persistent, searchable personal knowledge. Rather than wrapping a single model in a chat interface, Riera orchestrates specialized AI agents through explicit control flow and shared semantic memory.

Graph-Based Orchestration

User requests flow through a directed graph with distinct processing stages:

Intent Analysis

Specialist Execution

Response Synthesis

Intent Analysis

Specialist Execution

Response Synthesis

An LLM-powered router analyzes each request and produces a structured decision indicating which capabilities are required. Multiple specialists can execute in parallel for compound requests (e.g., "save this idea and remind me tomorrow" triggers memory storage and reminder scheduling simultaneously). Results are then synthesized into a unified response.

This separates routing logic from domain behavior — each component evolves independently.

Specialist Agents

Three agents handle distinct domains, each constrained by system prompts and capability boundaries:

Memory Specialist

Manages the full lifecycle of knowledge items: capturing new information, semantic retrieval, updates, and relationship management. Distinguishes between tasks (actionable), goals (aspirational), ideas (captured thoughts), and reminders (time-bound).

Reminder Specialist

Handles time-based operations including natural language temporal parsing ("next Tuesday at 3pm"), recurrence patterns, and scheduling management.

Conversation Specialist

Handles general dialogue that doesn't require persistent state changes.

All specialists share the same underlying LLM but operate with enforced role boundaries — the memory specialist cannot create reminders, and vice versa.

Semantic Memory

Information is stored with dense vector embeddings that encode meaning, not just keywords. When users search — even with different words than they originally used — the system finds items by semantic similarity.

This enables:

  • Approximate retrieval: "dentist thing" finds "root canal appointment with Dr. Smith"
  • Conceptual discovery: "health stuff" surfaces medical, fitness, and wellness items
  • Forgiveness for imprecise recall: Users don't need exact phrasing

Each stored item carries structured metadata: type classification, lifecycle status, priority, tags, relationships, and temporal information — enabling both semantic and structured queries.

Background Execution

Time-sensitive operations run asynchronously outside user sessions:

Reminder Delivery

A continuous background service polls for due reminders, sends mobile push notifications, tracks delivery status, and automatically schedules next occurrences for recurring reminders.

Knowledge Enrichment

A periodic service enhances stored goals and ideas with AI-generated summaries and contextual metadata, improving retrieval over time without blocking user interactions.

These services scale independently of real-time API traffic.

LLM Architecture

Models function as reasoning engines, not just text generators:

  • Structured decisions: The router outputs explicit decision objects, not free-form text
  • Executed operations: Agents invoke concrete operations that persist data and schedule actions — they don't just describe what they would do
  • Safety boundaries: All model interactions include safety configurations with graceful handling of refusals

Infrastructure & Privacy

Cloud-native deployment with separated services: real-time API handling, background reminder delivery, and background enrichment — each scaling based on its load characteristics.

Complete Data Isolation

Every database query is scoped by user. Agents only see data belonging to the current user. No cross-user data sharing, no collaborative features, no shared state.

Technical Differentiation

Real multi-agent coordination

Directed graph with routing, parallel execution, and synthesis

LLMs as reasoning engines

Structured decisions and executed operations, not just text generation

Semantic long-term memory

Vector embeddings for meaning-based retrieval

Asynchronous background agents

Time-sensitive operations independent of user sessions

Production infrastructure

Containerized services, managed databases, push notifications, horizontal scalability