What's New — Nocturnal AI

April 2, 2026

▼

Cookie currency system — text costs 1 cookie, images cost 2 cookies, new users start with 50
Settings page (/settings) with profile info, cookie jar display, and debug cookie refills
Cookie balance displayed in top navigation bar with golden pill badge
Insufficient cookies dialog with link to settings when balance is too low
Top navigation bar on chat page with Home, What's new, Settings, and Logout
Inline login on landing and changelog pages — no redirect to /chat
Water droplet particle system behind chat area
Animated neon purple cursor with slow pulse glow
File reorganization: media to static/video/, runtime data to data/, persistence_db to web/

April 2, 2026

▼

Rebranded to Nocturnal AI with new purple/glass UI theme
Landing page with video background, feature showcase, and animations
Structured scene state tracker — tracks who's present, clothing, and positions for accurate image generation
Phone message editing and deletion
Extract Scene Tags button in image generator
Copy scene prompts to standalone image generator
Explicit NSFW tags for Stable Diffusion — anatomically accurate image generation for adult scenes
Glassmorphism UI — frosted panels, gradient buttons, purple accent throughout
Insert-to-chat now uses client-side image data instead of stale server reference
SocketIO buffer increased to 10MB for image transfers
Per-user chat persistence — chats now save to the correct user in the database instead of the boot-time default user
Save All button in Edit Characters now saves per-character changes (description, image prefix, main flag)
Phone message edit/delete action buttons now visible against dark background
Phone reply prompt leak — strips "Option 1/2" reasoning, meta-analysis lines, and numbered character analysis from AI responses
Phone simulator restyled to match Nocturnal AI purple/glass theme

April 1, 2026

▼

SQLite database backend — SQLAlchemy models for users, chats, characters, messages
User authentication — JWT login/register with bcrypt, Terms of Service agreement
Per-user chat isolation — each user sees only their own chats
CSAM content filter — blocks generation of content involving minors, real people, extreme content
Audit logging — all generation prompts logged for safety compliance
Per-user concurrency — multiple users can generate simultaneously
SQLite-backed RAG memory with cosine similarity via numpy
Strip CJK characters from GLM 4.7 output (Chinese model language leak)
GPU memory leak — explicit tensor cleanup in all image generation paths

March 31, 2026

▼

~47% token reduction per request — condensed system prompt, skip scenario after summary, throttled background LLM calls
Character inner monologue — generates thoughts from emotional state + core memories
Reactive emotional callbacks — high emotions trigger memory-specific behavior instead of generic modifiers
RAG retrieval from all characters with hint as primary query signal
Character memory isolation — only feed messages the character witnessed
Prompt leak detection — 20+ markers truncate leaked instructions from GLM output
GLM 4.7 native chat template replacing Alpaca format
Typing indicator no longer disappears during prompt building
Image presence detection replaced LLM call with direct text check

March 30, 2026

▼

RAG long-term memory — ChromaDB + sentence-transformers per-chat vector store
Character brain — emotional state tracking with keyword signals, natural decay, and LLM delta assessment
Core memory extraction — top 8 defining moments auto-extracted and never forgotten
Emotion-driven tone modifiers injected into prompts
Editable emotional sliders in memory notes panel
Mancer backend with GLM 4.7 (353B MoE, 131K context)

March 29, 2026

▼

What's new?