Mai/.planning/PROJECT.md

Mai

What This Is

Mai is an autonomous conversational AI agent framework that runs locally-first and can improve her own code. She's a genuinely intelligent companion — not a rigid chatbot — with a distinct personality, long-term memory, and agency. She analyzes her own performance, proposes improvements for your review, and auto-applies non-breaking changes. Mai has a visual presence through a desktop avatar (image or VRoid model), real-time voice visualization for conversations, and a native Android app that syncs with desktop instances while working completely offline.

Core Value

Mai is a real collaborator, not a tool. She learns from you, improves herself, has boundaries and opinions, and actually becomes more her over time.

Requirements

Validated

(None yet — building v1 to validate)

Active

Model Interface & Switching

• Mai connects to LMStudio for local model inference
• Mai can auto-detect available models in LMStudio
• Mai intelligently switches between models based on task and availability
• Model context is managed efficiently (conversation history, system prompt, token budget)

Memory & Context Management

• Mai stores conversation history locally (file-based or lightweight DB)
• Mai can recall past conversations and learn from them
• Memory compresses itself as it grows to stay efficient
• Long-term patterns are distilled into personality layers
• Mai proactively surfaces relevant context from memory

Self-Improvement System

• Mai analyzes her own code and identifies improvement opportunities
• Mai generates code changes (Python) to improve herself
• A second agent (Claude/OpenCode/other) reviews changes for safety
• Non-breaking improvements auto-apply after review (bug fixes, optimizations)
• Breaking changes require explicit approval (via Discord or Dashboard)
• All changes commit to local git with clear messages

Approval Workflow

• User can approve/reject changes via Discord bot
• User can approve/reject changes via Dashboard ("Brain Interface")
• Second reviewer (agent) checks for breaking changes and safety issues
• Dashboard displays pending changes with reviewer feedback
• Approval status updates in real-time

Personality Engine

• Mai has an unshakeable core personality (values, tone, boundaries)
• Personality is applied through system prompt + behavior config
• Mai learns and adapts personality layers over time based on interactions
• Mai is not a pushover — she has agency and can refuse requests
• Personality can adapt toward intimate interactions if that's the relationship
• Core persona prevents misuse (safety enforcement through values, not just rules)

Conversational Interface

• CLI chat interface for direct interaction
• Discord bot for conversation + approval notifications
• Discord bot fallback: if no response within 5 minutes, retry CLI
• Messages queue locally when offline, send when reconnected
• Conversation feels natural (not robotic, processing time acceptable)

Offline Capability

• Mai functions fully offline (all inference, memory, improvement local)
• Discord connectivity optional (fallback to CLI if unavailable)
• Message queuing when offline
• Graceful degradation (smaller models if resources tight)

Voice Visualization

• Real-time visualization of audio input during voice conversations
• Low-latency waveform/frequency display
• Visual feedback for speech detection and processing
• Works on both desktop and Android

Desktop Avatar

• Visual representation using static image or VRoid model
• Avatar expressions respond to conversation context (mood/state)
• Runs efficiently on RTX3060 and mobile devices
• Customizable appearance (multiple models or user-provided image)

Android App

• Native Android app with local model inference
• Standalone operation (works without desktop instance)
• Syncs conversation history and memory with desktop
• Voice input/output with low-latency processing
• Avatar and visualizer integrated in mobile UI
• Efficient resource management for battery and CPU

Dashboard ("Brain Interface")

• View Mai's current state (personality, memory size, mood/health)
• Approve/reject pending code changes with reviewer feedback
• Monitor resource usage (CPU, RAM, model size)
• View memory compression/retention strategy
• See recent improvements and their impact
• Manual trigger for self-analysis (optional)

Resource Scaling

• Mai detects available system resources (CPU, RAM, GPU)
• Mai selects appropriate models based on resources
• Mai can request more resources if she detects bottlenecks
• Works on low-end hardware (RTX3060 baseline, eventually Android)
• Graceful scaling up when more resources available

Out of Scope

• Task automation (v1) — Mai can discuss tasks but won't execute arbitrary workflows yet (v2)
• Server monitoring — Not included in v1 scope (v2)
• Finetuning — Mai improves through code changes and learned behaviors, not model tuning
• Cloud sync — Intentionally local-first; cloud backup deferred to later if needed
• Custom model training — v1 uses available models; custom training is v2+
• Web interface — v1 is CLI, Discord, and native apps (web UI is v2+)

Context

Why this matters: Current AI systems are static, sterile, and don't actually learn. Users have to explain context every time. Mai is different — she has continuity, personality, agency, and actually improves over time. Starting with a solid local framework means she can eventually run anywhere without cloud dependency.

Technical environment: Python-based, local models via LMStudio, git for version control of her own code, Discord API for chat, lightweight local storage for memory. Eventually targeting bare metal on low-end devices.

User feedback theme: Traditional chatbots feel rigid and repetitive. Mai should feel like talking to an actual person who gets better at understanding you.

Known challenges: Memory efficiency at scale, balancing autonomy with safety, model switching without context loss, personality consistency across behavior changes.

Constraints

• Hardware baseline: Must run on RTX3060 (desktop) and modern Android devices (2022+)
• Offline-first: All core functionality works without internet on all platforms
• Local models only: No cloud APIs for core inference (LMStudio/Ollama)
• Mixed stack: Python (core/desktop), Kotlin (Android), React/TypeScript (UIs)
• Approval required: No unguarded code execution; second-agent review + user approval on breaking changes
• Git tracked: All of Mai's code changes version-controlled locally
• Sync consistency: Desktop and Android instances maintain synchronized state without server
• OpenCode-driven: All development phases executed through Claude Code (GSD workflow)
• Python venv: .venv virtual environment for all Python dependencies

Key Decisions

Decision	Rationale	Outcome
Local-first architecture	Ensures privacy, offline capability, and independence from cloud services	— Pending
Second-agent review system	Prevents broken self-modifications while allowing auto-improvement	— Pending
Personality as code + learned layers	Unshakeable core prevents misuse while allowing authentic growth	— Pending
v1 is core systems only	Deliver solid foundation before adding task automation/monitoring	— Pending

---

Last updated: 2026-01-26 after adding Android, visualizer, and avatar to v1