Ellibria v1.4.2.exe

Release Notes

Web Search

Ellibria can now search the web in real time using DuckDuckGo. Triggered automatically when your message contains relevant keywords — current events, prices, weather, people, dates. No API key required.

Fast / Expert Mode

New toggle in the top-right corner of the chat window. Fast mode keeps responses concise and snappy. Expert mode unlocks full depth — detailed analysis, step-by-step reasoning, no brevity constraints.

Auto-Update

The app now checks for new versions via GitHub Releases. A single click downloads and installs the update silently — the app restarts automatically with no reinstall required.

Photo Attachment

Attach images by clicking the photo icon or by dragging and dropping directly into the message input area. A thumbnail preview appears above the input with an × button to remove the image before sending.

Inner Thoughts Improvements

Ellibria's background thoughts and dream bar are now more authentic and grounded — no more self-referential doubts or technical self-analysis breaking the atmosphere.

Voice Recording Overlay Fix

The recording animation bubble now correctly adapts to any theme — light or dark — using dynamic color variables. The microphone icon stays visible and turns red during recording, returning to normal once done.

Token Efficiency

Significantly reduced token consumption per message — shorter context window, smarter profile caching, optimized liked message examples, and trimmed summary length. Faster responses and lower API usage.

Smarter User Profile

The profile engine now filters out situational noise and only records permanent facts — name, location, stable interests, personality traits, life goals. Identity-level facts are always preserved and never dropped.

Current Date Awareness

Ellibria now always knows today's exact date without needing to search for it, preventing date hallucinations in conversation.

Concurrency & Stability Fixes

Fixed a race condition in rolling summary generation, thread-safety issues with the dream counter, and potential crashes when image messages were passed through text-only processing pipelines.

Checking for updates

Added the ability to check GitHub releases for the new version. Auto-update was not included in the release because it was constantly breaking in the beta version due to Windows issues.

Full Changelog: v1.4.1...v1.4.2

Ellibria.v1.4.1.exe

🚀 Ellibria v1.4.1 — Advanced Voice HUD & Deep Architecture
In this update, we've elevated voice interaction to the level of fully-fledged modern interfaces, added strict language rules for models, and laid a strong foundation for working with local data and documents.

🎙️ Evolution of the Voice Interface (Voice HUD)
Interactive Voice Overlay: A new pop-up interface element (#voice-overlay) has been added—an elegant oval HUD that smoothly appears in the center of the screen when voice input is activated.

Live CSS Wave Animation: A visual indicator of audio capture (@Keyframes wave) has been integrated. Dynamic bars now pulse in real time, providing instant and clear feedback: the system is listening to you.

Themed Adaptability of the Voice Panel: The overlay logic has been completely converted to system CSS variables (var(--bg2), var(--green), var(--border)). The recording panel perfectly adapts to any selected theme (from Dark Minimal to Clean Light), preserving soft shadows and avoiding unnecessary visual contrast.

Safe Recording Lifecycle: The onstart, onerror, and onend triggers have been redesigned in main.js. The overlay is guaranteed to hide upon any microphone failure or speech completion, eliminating the possibility of the interface freezing in "listening" mode.

🧠 Fixing the Language Behavior of Models
Protection against Linguistic "Hallucinations": A strict language control rule has been implemented in the system instructions (MODE_PROMPTS) for all modes (Default, Dommy, Hacker, etc.). Models are strictly prohibited from mixing languages ​​within a single response or using random characters (Japanese/Chinese scripts) unless the user has explicitly initiated the conversation.

⚙️ Backend Architecture Enhancements and Stability
Critical Time Error Fix: In app.py, a bug that caused a server crash due to a missing import of the time module has been localized and fixed. This has completely stabilized asynchronous background features, such as the delayed_dream function.

Preparing for Database Integration: The sqlite3 module has been imported into the backend core. This lays the architectural foundation for the transition from custom JSON files to a full-fledged, fast, and secure local database for storing conversation history.

Native Web Resource Calls: The webbrowser module has been integrated, allowing the application to correctly and securely open external links in the user's default browser directly from the desktop window context.

📂 New Data Horizons
PDF Parser Integration: The pypdf library has been added to the dependency stack (requirements.txt). Ellibria is now ready to implement features for reading, parsing, and indexing local PDF documents for RAG.

Full Changelog: v1.4.0...v1.4.1

Ellibria v1.4.0.exe

🚀 Ellibria v1.4.0 — Native UX & Context Stability
In this update, we transformed Ellibria into a full-fledged Windows desktop application and significantly enhanced stability for both cloud and local workflows.

🌟 Native Desktop Experience
System-Aware Theming: Ellibria now detects your Windows system-wide light/dark mode settings on launch and automatically applies the corresponding theme for a seamless look.

Native Data Export: Clicking “Export JSON” now triggers the standard Windows “Save As...” dialog, allowing you to choose your backup location with ease.

Secure Path Handling: All user-specific data (logs, configurations, keys) is now strictly isolated within the %USERPROFILE%/.ellibria-agent directory, ensuring full compliance with Windows security policies.

🎙️ Enhanced Voice Interaction (STT)
Integrated Speech-to-Text: Added native voice input capabilities via the Web Speech API.

Adaptive Recognition: Automatically syncs with your interface language (e.g., en-US, ru-RU) for accurate hands-free interaction.

Visual Feedback: The microphone button now features status-based pulsing (recording), and the system displays “Listening...” toasts to keep you informed.

⚙️ Performance & Stability Polish
Context Consistency Engine (Anti-Crash): The history cleanup logic now deletes messages strictly in pairs (user/assistant), preventing 400 Bad Request errors during extended sessions.

Smart Memory Capping: Local LLM memory protection is now active. The background thread automatically truncates old context to ~1,500 characters, eliminating VRAM overflow risks during long-running sessions.

Robust JSON Parsing: A protective layer now automatically strips Markdown code blocks (e.g., json ...) from local model responses, preventing parser crashes.

Vision Error Guard: Improved error handling for Vision models. API rate limits (e.g., 429 Too Many Requests) are now caught gracefully, ensuring the generation pipeline remains intact.

🔧 UI/UX Refinements
Dynamic Configuration Interface: Completely redesigned the provider switching logic in setup_and_run.py.

Smart UI Anchoring: The API key input field now intelligently appears/disappears based on the selected provider.

Layout Stability: By using dynamic UI anchoring (before=btn), we ensured the configuration layout remains perfectly aligned, eliminating visual "ghosting" or element displacement when toggling between Cloud and Local models.

Full Changelog: v1.3.0...v1.4.0

Ellibria v1.3.0.exe

🚀 Ellibria v1.3.0 — Dynamic Memory, Safe Mode & UI Polish
This update focuses on making conversations with Ellibria much deeper and more natural, while adding smart hardware adaptation to protect weak GPUs and introducing a personalized tone-mirroring system via liked messages.

🌟 New Features
Dynamic Context Depth: The agent now automatically scales its memory based on the active backend to maximize the depth of your dialogue:

Gemini Flash: Remembers up to 40 messages for an incredibly rich, long-term conversation experience.

Groq API: Keeps 20 messages in context, balancing ultra-fast response times with a coherent dialogue flow.

Smart Safe Mode (Toggle for Weak GPUs): Added a dedicated "Safe Mode" switcher in the Settings panel designed as a "safe default" for the app.

When ON (Default): It automatically analyzes the size of your local model. Heavy models (e.g., Llama 3 8B, Mistral) are strictly capped to 6 messages to prevent VRAM overflow and desktop freezes on cards like GTX 1060. Lighter models (e.g., Phi-3, Gemma 2B) automatically expand to 12 messages.

When OFF: It completely removes constraints and unlocks the full base memory cache (15+ messages) for high-end desktop setups (e.g., RTX 4090).

Personalized Style Mirroring (Heart Button): Liking a message now actively shapes Ellibria's personality.

Pressing the heart button saves responses to a dedicated cache (up to 10 entries with auto-rotation).

Saved responses are injected into the system prompt as stylistic examples, allowing the model to gradually mirror the tone, length, and formatting you prefer.

Full unlike/re-like toggle logic supported.

📈 Improvements & UI Polish
Precise Search Navigation: Searching no longer awkwardly jumps to the bottom of the screen. It now scrolls directly to the matched message bubble and flashes it with a elegant 2-second accent outline.

Auto-Port Allocation: Added find_free_port(). The application automatically scans and binds to a free port in the 5000–5010 range on startup, preventing launch failures if port 5000 is occupied by another process.

Fully Offline-Capable: All core CDN dependencies (marked.js, highlight.js, github-dark.css) have been mirrored into local static directories.

Themed Code Blocks: Code syntax blocks no longer suffer from a hardcoded dark theme background (#0d1117) — they now dynamically inherit var(--bg3) and respect your selected theme.

UI Toggle Switch: Added a sleek, minimalist animated toggle switch for Safe Mode that perfectly morphs into your selected visual theme.

🐛 Bug Fixes
Surgical Session Deletion: Fixed delete_session wiping the entire global memory file. It now surgically targets and removes only the rolling summary of the deleted chat session, preserving all other context.

Resilient Rolling Summary: Fixed a silent crash (AttributeError) in background memory generation when the local AI engine temporarily failed to initialize (added a client null-check guard).

Persistent Sessions: Fixed Flask secret_key regenerating on every software restart. Keys are now safely persisted to a .secret_key file on the first launch, preventing user sessions from constantly invalidating.

Animation Fix: Resolved a missing @Keyframes floatAvatar definition; the avatar floating animation now triggers properly.

Clean Code Structure: Fixed an indentation anomaly for comment # 5. inside the chat() function.

Adjusted the User Personality Archive logic — the model now explicitly prioritizes fresh, in-conversation constraints over archived historical facts if an explicit contradiction arises.

🔧 Under the Hood & Technical Highlights
Thread Safety Multiplier: Synchronized background memory compression (Rolling Summary) and the Profile Analyzer using explicit RLock file locking (with file_lock:), completely mitigating JSON data corruption risks during parallel writes.

Launch Chain Overhaul: Port bindings are now gracefully channeled through the full lifecycle: find_free_port() → start_flask() → open_webview().

Added explicit hiddenimports to Ellibria.spec to guarantee PyInstaller binary compilations don't throw unexpected ModuleNotFoundError crashes at runtime.

Fail-safe string parsing wrappers added for connected backend engines and model names.

Full Changelog: v1.2.1...v1.3.0

Ellibria v 1.2.1.exe

Update Notes:

• Added “Profile Archive” section. This is the core of your identity, and if the AI agent remembered something incorrectly or missed something, you can add or remove a fact about yourself

• Added a button to import and export the Profile Archive. When switching to another device, you can simply copy the contents of the folder C:\Users<YOU>.ellibria-agent

• When the sidebar is minimized, there are no longer any unsightly icons for existing chats. For a minimalist look, chat icons have been completely removed from the minimized version of the sidebar

• Mode names have been changed and optimized

• When you change the theme in Windows, the app’s theme changes as well. However, if you changed the theme upon opening the app, the app’s theme will not revert when you change the Windows theme

• Added a notification when a new fact is added to the Personal Archive

• Added error logging to the C:\Users<YOU>.ellibria-agent folder in the error.log file. This is very convenient and lets you know when the AI agent encounters errors

• Added an animated “Agent is typing” indicator

• Added smooth scrolling and more elegant text scaling when the window expands

• Redesigned the dialog box that opens when deleting a chat

• Completely removed all references to the former name, Echo

Full Changelog: v1.2...v1.2.1

Ellibria.1.2.exe

Release Notes:

Major Changes and Improvements

In this update, we have significantly expanded the agent’s memory capabilities and improved the application’s stability.

- Introduction of the “Personality Archive” (Core Personality Engine):

The agent now possesses dynamic long-term memory. It automatically analyzes dialogues, extracts key facts about the user (preferences, communication style, important data), and saves them to a personal “Personality Archive.”

This data is now factored into the system prompt, allowing Ellibria to be more personalized and attentive to detail in long-term conversations.

- Full Data Management (Profile Management):

New API endpoints have been added for managing the archive. You can now view, edit, or manually add facts about yourself through the interface.

- Multithreaded Architecture (Performance Boost):

Personality analysis and data saving now occur in background threads. This has completely eliminated AI response delays—the chat runs smoothly while the background agent populates the knowledge base.

- Enhanced Reliability (Data Safety):

A locking system (RLock) has been implemented to prevent any access conflicts when writing to JSON files (settings, memory, sessions). Your chat history is now protected against data loss.

- Interface and UX Improvements:

The window display mechanism has been completely redesigned. Layout conflicts between the history search and the profile overlay have been fixed.

Visual loading indicators (spinners) and input locking while processing a request have been added.

Textarea behavior has been optimized (automatic resizing is now smoother).

Fixes

Fixed a critical HTML nesting error that caused the UI to disappear during searches.

Eliminated the possibility of server crashes due to invalid incoming JSON requests.

Improved encoding handling to ensure Cyrillic support in all configuration files.

Full Changelog: v1.0...v1.2

Ellibria.1.0.exe

Ellibria v1.0 — First Public Release

Windows installer. No Python required — just download, run, and enter your Groq API key.

Get your free API key at console.groq.com/keys

What's included:

• Multiple personality modes (Mistress, Mr. Robot, Psychotherapist, Loyal Friend and more)
• Persistent memory between sessions
• Voice output with female TTS
• Dark, Light, Purple, Night and Terminal themes
• Auto-theme sync with Windows system theme
• Smart backend detection (LM Studio, Ollama, Groq, Gemini)
• Fully customizable system prompt

Installation:

1. Download Ellibria_Installer.exe
2. Run the installer
3. Launch Ellibria and enter your Groq API key
4. Done

Feedback and bug reports welcome via Discord or email.