Aura Backend - Advanced AI Companion
Sophisticated AI Companion with Vector Database, Emotional Intelligence, and Model Context Protocol Integration
Current emotion assessment behavior
Aura now distinguishes tentative user-emotion interpretations, its own simulated tone, deliberate abstention, invalid model output, and unavailable analysis. Each accepted proposal includes checked verbatim source quotations and the hash of the analyzed text. Failed analysis stays Unknown, including in stored history; it no longer silently becomes “Normal.” Text does not measure a user's brain activity or chemical levels. The interface labels Aura's indicators as simulations derived from saved controller state; an unknown classification does not erase those readouts.
The local Ornith smoke evaluation handled 11 of 12 invented cases as expected; it incorrectly labeled an ambiguous sarcastic sentence. Its strict acceptance gate therefore failed. This is a tested validation boundary, not evidence of reliable general emotion recognition. See the implementation and evaluation report for the exact behavior, limitations, and rerun commands. Historical data has not been rewritten.
Two WARNINGS and a disclaimer
-
AI generated code
-
Aura could be dangerous despite my attempted safeguards in a number of ways including but not limited to PC damage User mental health and attachment Emotional agentic activity
User assumes all liability.
🌟 Features
🧠 Advanced Cognitive Architecture
- ASEKE Framework: Adaptive Socio-Emotional Knowledge Ecosystem
- Tentative Emotion Assessment with checked source quotations and explicit uncertainty
- Cognitive Focus Tracking across different mental frameworks
- Adaptive Self-Reflection for continuous improvement
- 🆕 Thinking Extraction: Transparent AI reasoning with thought analysis and cognitive transparency
🗄️ Intelligent Memory System
- Vector Database Integration with ChromaDB for semantic search
- Persistent Conversation Memory with embedding-based retrieval
- Emotional Pattern Analysis over time
- Cognitive State Tracking and trend analysis
- MemVid AI QR code mp4 memory Infinite MP4 based memory
- Internal AI guided Memory Organization tools Move information from short to long term memory systems to avoid bottlenecks and categorize chats
🔗 MCP Integration
- Model Context Client Utilizes the same MCP config JSON format as Claude Desktop- Use ANY tools!
- Model Context Protocol Server for external tool integration
- Standardized AI Agent Communication following MCP specifications
- Tool Ecosystem Compatibility with other MCP-enabled systems
- Bidirectional Data Exchange with external AI agents
📊 Advanced Analytics
- Emotional Trend Analysis with stability metrics
- Cognitive Pattern Recognition and optimization
- Personalized Recommendations based on interaction history
- Data Export in multiple formats (JSON, CSV, etc.)
Data Flow
- User Input → Frontend → FastAPI
- Processing → Vector DB Search → Context Retrieval
- AI Processing → Explicitly Selected Provider → Response Generation
- State Updates → Emotional/Cognitive Analysis → Pattern Storage
- Memory Storage → Vector DB → Persistent Learning
- External Access → MCP Server → Tool Integration
🧠 AI Thinking & Reasoning Transparency
Thinking Extraction Capabilities
- Real-time Reasoning Capture: Extract and analyze AI thought processes during conversations
- Thought Summarization: Automatic generation of reasoning summaries for quick understanding
- Cognitive Transparency: Full visibility into how Aura approaches problems and makes decisions
- Reasoning Metrics: Detailed analytics on thinking patterns, processing time, and cognitive load
Thinking Configuration
- Thinking Budget: Configurable reasoning depth (1024-32768 tokens)
- Response Integration: Optional inclusion of reasoning in user responses
- Pattern Analysis: Long-term analysis of reasoning patterns and cognitive development
- Performance Optimization: Thinking efficiency metrics and optimization recommendations
🎭 Emotional Intelligence System
Simulated Emotions
The header names the strongest tendency in Aura's committed controller state: Calm, Curious, Excited, Concerned, Warm, Content, or Peaceful. Conversation events are proposed by the selected model, checked against exact source quotes, and translated into bounded, authored state changes before the reply is generated. User-emotion inference and Aura's optional reply-style analysis are separate records; a failed reply analysis cannot freeze or overwrite the controller.
Simulated Indicators
- Brainwave labels: Alpha, Beta, Gamma, Theta, Delta are activation bands; the percentage exposes changes within a band.
- Chemical channels: Six controller readouts appear in the simulation panel. The header shows the largest channel, which can remain stable while others change.
- State persists with the conversation and decays toward baseline between turns. These are software analogies, not measured EEG or chemical concentrations. User assessments carry no biological labels.
See the conversation simulation repair for the causal path, checks, and limitations.
🧠 ASEKE Cognitive Framework
Components
- KS (Knowledge Substrate): Shared conversational context
- CE (Cognitive Energy): Mental effort and focus allocation
- IS (Information Structures): Ideas and concept patterns
- KI (Knowledge Integration): Learning and connection processes
- KP (Knowledge Propagation): Information sharing mechanisms
- ESA (Emotional State Algorithms): Emotional influence on processing
- SDA (Sociobiological Drives): Social dynamics and trust factors
📊 Analytics & Insights
Emotional Analysis
- Stability Metrics: Emotional consistency over time
- Dominant Patterns: Most frequent emotional states
- Transition Analysis: Emotional state changes and triggers
- Intensity Tracking: Emotional intensity distribution
- Brainwave Correlation: Neural activity pattern analysis
Cognitive Tracking
- Focus Patterns: ASEKE component utilization
- Learning Efficiency: Knowledge integration rates
- Context Switching: Cognitive flexibility metrics
- Attention Allocation: Cognitive energy distribution
🚦 Performance-
Responses take some time to process depending on tasks, any coder wants to see if they can speed up the processes I would be grateful.
Optimization
- Vector database indexing for fast searches
- Async processing for concurrent requests
- Cost-Free Local Embeddings: Support for
Ollamaandfastembed(BGE/Gemma) to avoid API costs - Autonomous sub-model background Focus gating and task processing for state updates and tool use
- Tool learning adapter
- MemVid Infinite memory with modern
.mv2single-file archival!
Monitoring
- Health check endpoint
- Performance metrics collection
- Error tracking and reporting
- Resource usage monitoring
MCP Client now fully functional!!! Memvid integration attempted- still testing.
I am not a coder so hopefully it sets up right if anyone tries it.
Supported local startup
Aura is a private, single-user local application. It has no sign-in layer and
binds to 127.0.0.1 by default. Run these commands from the repository root.
One-time dependency setup
Setup is an explicit operator action. The startup command and wrapper scripts do not install, synchronize, or download software or models.
uv sync --locked
To enable the real Memvid archive integration, install the locked extra instead:
uv sync --locked --extra memvid
Set AURA_MEMVID_ENABLED=true and select MEMVID_EMBEDDING_PROVIDER=ollama
with your installed MEMVID_EMBEDDING_MODEL (for example embeddinggemma:latest).
Set MEMVID_TELEMETRY=0 to disable SDK analytics. This adapter uses precomputed
local vectors, not the SDK's implicit cloud embedding selection.
Memory storage has three distinct roles: SQLite keeps committed conversations;
Chroma indexes embeddings for active semantic search; Memvid keeps independent
.mv2 archive snapshots with their own vectors. In the UI, Archive this chat
copies the latest 100 exchanges and verifies the saved content after reopening.
It does not delete active messages or import old Chroma/video archives. Archive
files live below the configured ledger directory in memvid/. Search checkboxes
select active memory, archives, or both. Aura also receives archive_session
and search_archives tools when Memvid starts successfully.
The opt-in real local smoke test uses temporary synthetic data and tests chat continuity, archival, search, and an application restart:
uv run --locked --no-sync python scripts/verify_local_memory.py
The maintained Aura Modelfile is docs/models/ornith-apex/Modelfile.aura.
Rebuild with ollama create aura-ornith:35b -f docs/models/ornith-apex/Modelfile.aura.
It requests a 131072-token context and an 8192-token generation budget; actual
generation can use a smaller explicit request budget. Context allocation does
not guarantee accurate recall at the full capacity. The locally tested recall
probe used 30000 input tokens. AURA_HISTORY_MAX_CHARS is a separate application
history budget in characters, not tokens; keep it below the model's capacity.
npm ci
Copy .env.example to .env only if you want to customize the local defaults.
The example selects Ollama and contains no credential. Gemini and OpenRouter are
optional cloud providers and require an explicit provider selection plus the
corresponding credential in your private environment.
Preflight, then serve
For normal use, run the serve command below; it runs preflight automatically.
On Linux, the existing launcher is the shorter equivalent: ./start_full_system.sh.
Both commands load this repository's .env without extra flags. Exported shell
variables take precedence. AURA_MODEL works with any selected provider;
OPENROUTER_MODEL or OLLAMA_MODEL, when set, takes precedence over it.
Selecting OpenRouter does not require a local Ollama chat model.
Active memory search uses AURA_EMBEDDING_PROVIDER and AURA_EMBEDDING_MODEL.
The compatibility default is sentence_transformers / all-MiniLM-L6-v2.
For local Ollama embeddings, select ollama / embeddinggemma:latest;
AURA_EMBEDDING_BASE_URL optionally overrides OLLAMA_BASE_URL for embeddings.
The separate MEMVID_EMBEDDING_* settings apply only to optional archives.
Changing the active embedding model builds and verifies a new derived index on
first use before switching; the old generation and SQLite records are retained.
Failed embedding requests leave the rebuild incomplete rather than switching to
an invalid index. This may make the first memory operation slower.
Conversation continuity uses committed exchanges from the same user/session,
not just a provider session identifier. It retains up to 100 recent exchanges
within AURA_HISTORY_MAX_CHARS (default 24000 characters, not tokens). Increase
this only alongside a verified model context allocation. Chat history uses real
session IDs and stable first-message titles; its list requests are capped at 100.
Startup initializes and opens the selected ledger before reporting readiness.
AURA_CLEAN_INSTALL=true selects the new SQLite-backed read path without importing
legacy stores. Chroma remains the derived semantic index; Memvid is a separate
optional archive integration, not a prerequisite for chat.
uv run --locked --no-sync python -m aura_backend.runtime preflight
Preflight is report-only. It checks Python, uv, Node, npm, both lock contracts, provider configuration, the selected port and storage paths, the selected provider service and selected model, and application readiness. The provider rows are a bounded live provider check; they are not part of the offline test suite. Preflight never installs dependencies, downloads a model, creates storage, changes permissions, kills another process, or starts Aura.
The JSON status is one of pass (exit 0), missing (2), failed (3),
blocked (4), not_run (5), or not_applicable (6). Only a complete pass
licenses startup. Other results name a safe remediation code; perform any repair
explicitly and rerun preflight rather than treating a blocked check as readiness.
uv run --locked --no-sync python -m aura_backend.runtime serve
serve runs preflight first, starts only the requested local child processes,
waits for the backend /ready response, and returns a nonzero status if startup
or a child fails. Ctrl+C/SIGTERM cleans up only processes and local provider
sessions owned by this invocation. Cancellation cannot guarantee stopped remote
compute or billing at a cloud provider.
The cross-platform launchers are thin delegates to these same commands:
./start_full_system.sh and start_full_system.bat run full serve;
./aura_backend/start_api.sh and ./aura_backend/start_frontend.sh select one
side. ./aura_backend/start_mcp.sh is a separate optional MCP delegate and is
not part of normal Aura readiness.
For normal private use, keep the loopback default. Passing a non-loopback
--host is explicit LAN exposure; the runtime warns that Aura has no sign-in.
Do not expose Aura directly to the internet.
Once serve reports readiness, the local UI is at http://localhost:5173, the API at http://localhost:8000, and API documentation at http://localhost:8000/docs. Environment-blocked and live-provider results are evidence about that machine only, not proof that every provider or model works.
For current background processing, memory behavior, configuration changes and verification, see Autonomic work and memory.
📡 API Endpoints
Core API
- Health Check:
GET /health - Process Conversation:
POST /conversation - Search Memories:
POST /search - Emotional Analysis:
GET /emotional-analysis/{user_id} - Export Data:
POST /export/{user_id}
API Documentation
Visit http://localhost:8000/docs for interactive API documentation.
🔗 MCP Integration
Available MCP Tools- Working on emotional state records, hopefully fixed tomorrow
- search_aura_memories: Semantic search through conversation history
- analyze_aura_emotional_patterns: Deep emotional trend analysis
- store_aura_conversation: Add memories to Aura's knowledge base
- get_aura_user_profile: Retrieve user personalization data
- export_aura_user_data: Data export functionality
- query_aura_emotional_states: Information about emotional intelligence system
- query_aura_aseke_framework: ASEKE cognitive architecture details
Connecting External Tools
To connect external MCP clients to Aura:
Example MCP client configuration- for Claude or other clients to talk to Aura or use as a system.
Edit your directory path and place in claude desktop config json.
{
"mcpServers": {
"aura-companion": {
"command": "uv",
"args": [
"--directory",
"/home/ty/Repositories/ai_workspace/emotion_ai/aura_backend",
"run",
"aura_server.py"
]
}
}
}
🏗️ Architecture
System Components
┌─────────────────────────────────────────────────┐
│ Frontend │
│ (React/TypeScript) │
└─────────────────┬───────────────────────────────┘
│ HTTP/WebSocket
┌─────────────────▼───────────────────────────────┐
│ FastAPI │
│ (REST API Layer) │
├─────────────────┬───────────────────────────────┤
│ │ │
│ ┌──────────────▼─────────────┐ │
│ │ Vector Database │ │
│ │ (ChromaDB) │ │
│ │ │ │
│ │ • Conversation Memory │ │
│ │ • Emotional Patterns │ │
│ │ • Cognitive States │ │
│ │ • Knowledge Substrate │ │
│ └────────────────────────────┘ │
│ │
│ ┌────────────────────────────┐ │
│ │ State Manager │ │
│ │ │ │
│ │ • Emotional Transitions │ │
│ │ • Cognitive Focus Changes │ │
│ │ • Automated DB Operations │ │
│ │ • Pattern Recognition │ │
│ └────────────────────────────┘ │
│ │
│ ┌────────────────────────────┐ │
│ │ File System │ │
│ │ │ │
│ │ • User Profiles │ │
│ │ • Data Exports │ │
│ │ • Session Storage │ │
│ │ • Backup Management │ │
│ └────────────────────────────┘ │
└─────────────────┬───────────────────────────────┘
│ MCP Protocol
┌─────────────────▼───────────────────────────────┐
│ MCP Server │
│ (External Tool Access) │
│ │
│ • Memory Search Tools │
│ • Emotional Analysis Tools │
│ • Data Export Tools │
│ • ASEKE Framework Access │
└─────────────────────────────────────────────────┘
🧪 Testing
Health Check (Working)
curl http://localhost:8000/health
Thinking Functionality Tests (New!)
# Test thinking extraction capabilities
cd aura_backend
python test_thinking.py
# Interactive thinking demonstration
python thinking_demo.py
# Check thinking system status
curl http://localhost:8000/thinking-status
Unit Tests
pytest tests/
Integration Tests
./test_setup.py
Load Testing
# Example using wrk
wrk -t12 -c400 -d30s http://localhost:8000/health
Local Development
I apologize for the mess, I do not know if any of this works below but feel free to try if you are brave or know what you are doing.
Production (Docker)
# Build image
docker build -t aura-backend .
# Run container
docker run -p 8000:8000 -v ./aura_data:/app/aura_data aura-backend
Systemd Service
# Copy service file
sudo cp aura-backend.service /etc/systemd/system/
# Enable and start
sudo systemctl enable aura-backend
sudo systemctl start aura-backend
🤝 Integration with Frontend
API Endpoints to Update
Update your frontend to use these endpoints:
const API_BASE = "http://localhost:8000";
// Replace localStorage with API calls
const response = await fetch(`${API_BASE}/conversation`, {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
user_id: userId,
message: userMessage,
session_id: sessionId,
}),
});
WebSocket Support (Future)
Real-time updates and streaming responses will be available via WebSocket connections.
📚 Advanced Usage
Custom MCP Tools
Create custom MCP tools by extending the mcp_server.py:
@tool
async def custom_aura_tool(params: CustomParams) -> Dict[str, Any]:
"""Your custom tool implementation"""
# Implementation here
pass
Vector Database Queries
Direct vector database access for advanced queries:
from main import vector_db
results = await vector_db.search_conversations(
query="emotional support",
user_id="user123",
n_results=10
)
🐛 Troubleshooting
Use the safe remediation codes in the startup guide. Aura never kills an unknown process, deletes a database, prints a credential, or rebuilds an environment as part of troubleshooting. Storage diagnosis and repair remain preservation-gated work; make a verified backup before any manual change.
Logs
Check logs in:
- Console output during development
- System logs:
journalctl -u aura-backend(if using systemd) - Application logs:
./aura_data/logs/
🔒 Security- WARNING! AI Generated so I have 0 trust in these features
Data Protection
- All user data stored locally
- Local Ollama keeps model traffic local; explicitly selected cloud providers transmit requests under their own terms
- Embeddings and archives remain local data and can retain sensitive information
- The default local HTTP connection is not encrypted
Access Control
- No sign-in or API authentication; keep the default loopback boundary
- Rate limiting enabled
- CORS configuration
- Input validation and sanitization
🛣️ Roadmap
Upcoming Features
- Real-time WebSocket connections
- Advanced emotion prediction models
- Multi-user collaboration features
- Enhanced MCP tool ecosystem
- Mobile app backend support
- Advanced analytics dashboard
- Integration with external AI models
Long-term Vision
- Multi-modal interaction (voice, video, text)
- Federated learning across Aura instances
- Advanced personality adaptation
- Enterprise deployment options
- Open-source community ecosystem
📄 License
My stuff is MIT I suppose but there is other software like google-genai and memvid so it is a mixed bag I think ie don't steal my ideas and try to make money, without me. lol but I am super poor.
🤝 Contributing
Contributions welcome! Please read our contributing guidelines and submit pull requests for review.
📞 Support
For issues and support:
- Check troubleshooting section
- Review logs and error messages
- Create detailed issue reports
- Join community discussions
Aura Emotion AI - Powering the future of AI companionship and assistance through advanced emotional intelligence and sophisticated memory systems.
