AZUS AI
About the project
AZUS is an intelligent personal assistant ecosystem that solves the context window problem in LLM-based chat. Instead of passing raw conversation history into every API call, AZUS uses PDM (Pressure-Driven Memory) — automatically extracting key facts, compressing them into signed memory signatures, and dynamically injecting them across sessions without exhausting token limits.
AZUS AI
About the project
AZUS is an intelligent personal assistant ecosystem that solves the context window problem in LLM-based chat. Instead of passing raw conversation history into every API call, AZUS uses PDM (Pressure-Driven Memory) — automatically extracting key facts, compressing them into signed memory signatures, and dynamically injecting them across sessions without exhausting token limits.
Project Overview
AZUS is a three-part ecosystem: a React-based Companion App for smart chat, a Python/Django backend housing the PDM memory engine, DocVault document storage, accounts and subscriptions, and Junior client endpoints, and a Junior Desktop App for local workflow integration. The platform's core differentiator is PDM — a Pressure-Driven Memory system that eliminates the context window bottleneck common to all LLM-based chat products.
Project Goal
The goal was to build an intelligent personal assistant ecosystem that solves the context window problem inherent in LLM-based chat. Standard products pass the entire conversation history into every API call — this fills the context window quickly and becomes prohibitively expensive. AZUS replaces this with PDM: the system automatically extracts important facts from conversations (preferences, plans, passwords, deadlines), compresses them into short signed memory signatures, and dynamically injects them into future prompts — even across completely separate chat sessions.
Solution
We built the complete AZUS platform: a Django backend with isolated apps for PDM memory management, DocVault document ingestion, Stripe-based subscriptions, and Junior Desktop endpoints; a React Companion App with Grok (X.AI) as the primary LLM and Voyage AI for 1024-dimensional semantic embeddings; a Celery + Redis async pipeline for background fact extraction and memory audits; OpenAI Whisper for cloud voice transcription and Vosk for offline local recognition; and a PDM-T temporal layer that computes deadline pressure and auto-surfaces urgent memories via TAS bypass when E_T exceeds 0.75.
Key Challenges
Standard LLM chat passes raw message history into every API call — context windows fill quickly, costs scale uncontrollably, and conversational coherence breaks down in long sessions
Users expect the assistant to remember preferences, plans, and deadlines across completely separate chat sessions — something raw message history cannot support
Time-sensitive information such as deadlines and appointments needs to automatically escalate in priority as the moment approaches — static memory systems cannot model temporal pressure
Memory extraction, semantic re-ranking, and background goal audits must run asynchronously without blocking chat response latency
Key Features
PDM (Pressure-Driven Memory) — Automatically extracts key facts from conversations, compresses them into signed memory signatures with p_magnitude, t_persistence, and phase_privilege parameters, and injects them dynamically into future prompts.
Cross-Chat Recall — Memory signatures persist across all separate chat sessions — the assistant always knows your active goals, preferences, and pending tasks regardless of which conversation they were mentioned in.
PDM-T Temporal Memory — When a memory contains a deadline or appointment, the system computes temporal curvature (c_temporal), membrane thickness (s_temporal), and temporal pressure (p_temporal). When E_T > 0.75, the signature is marked urgent and surfaced via TAS bypass.
DocVault — Automatically reads user-uploaded PDF and text documents, extracts summaries, and stores them in the PDM — making document knowledge queryable through natural language chat.
Junior Desktop Client — A native desktop companion with offline voice recognition via Vosk, local workflow integration, and dedicated Junior API endpoints on the Django backend.
Final Interface
Features Overview
Intelligent Personal Assistant Ecosystem
Three components — Companion App, Django Backend, and Junior Desktop — powered by PDM: a memory engine that extracts key facts, compresses them into signatures, and injects them across sessions to eliminate context window limits.
Tech Stack
Built on Python 3.13 with Django and Django REST Framework for the backend, PostgreSQL with pgvector for 1024-dimensional vector memory storage, Celery and Redis for async background processing, Grok (X.AI) as the primary LLM, Voyage AI for semantic embeddings, OpenAI Whisper for voice transcription, and AWS S3 for document storage.
postgresql + pgvector
cloud infrastructure
Metrics
0
Ecosystem components
Companion App, Django Backend, and Junior Desktop Client — three tightly integrated layers forming a unified intelligent assistant experience built around the PDM memory engine.
0000
Memory vector dimensions
Each PDM memory signature is stored as a 1024-dimensional pgvector embedding, enabling high-precision cosine similarity search across the full memory graph via direct SQL queries.
PDM-T
Temporal memory engine
When a memory contains a deadline or appointment, the system computes c_temporal, s_temporal, and p_temporal — auto-escalating urgent facts via TAS bypass when temporal energy E_T exceeds 0.75.
AI
Cross-Chat Recall
Memory signatures persist across all separate chat sessions — the assistant always knows your preferences, active goals, and pending deadlines regardless of which conversation they were mentioned in.
Companies That Build With Us
By clicking "Accept" you agree with our