What if AI could not only access memories, but consciously choose what to remember? With MCP tool access fully supported?
AI conversation platform implementing dual-layer memory architecture inspired by human cognition. Combines automatic background memory with conscious, deliberate memory operations that AI controls. Tool access powers similar to Claude Desktop.
Purpose-driven autonomous execution replacing simple query generation with sophisticated multi-step workflows:
- Purpose analysis and strategic planning
- Context gathering from all memory systems
- Smart tool orchestration with error recovery
- Progress monitoring with adaptive replanning
- Reflection engine for continuous learning
- Configurable aggressiveness and safety controls
Automatic Memory (RAG): Non-volitional background memory using ChromaDB vectors and Google text-embedding-004
Conscious Memory: Volitional operations via MCP tools - save, search, update, delete with tags and importance scoring
Knowledge Graph: Neo4j-powered relationship mapping with automatic synchronization and retry mechanisms
Exposes memory operations as Model Context Protocol tools for natural conversation flow. Clean separation between UI, memory, and AI operations.
- Node.js 18+
- Docker & Docker Compose
- LLM API key (free Google AI Studio recommended)
git clone https://github.com/esinecan/skynet-agent.git
cd skynet-agent
npm install
cp .env.example .env.local
# Edit .env.local with your API keys
docker-compose up -d # ChromaDB (8000) + Neo4j (7474, 7687)
npm run dev # Or npm run dev:next if Neo4j issuesAccess:
- Application:
http://localhost:3000 - Conscious Memory:
http://localhost:3000/conscious-memory - Neo4j Browser:
http://localhost:7474(neo4j/password123)
| Provider | Best For | Model |
|---|---|---|
| Multimodal & speed | gemini-2.5-flash-preview-05-20 |
|
| DeepSeek | Cost-effective | deepseek-chat |
| OpenAI | Ecosystem | gpt-4o-mini |
| Anthropic | Reasoning | claude-3-5-haiku-20241022 |
| Groq | Ultra-fast | llama-3.3-70b-versatile |
| Mistral | Natural language | mistral-large-latest |
| Ollama | Privacy | llama3.2:latest |
# LLM (pick one)
GOOGLE_API_KEY=your_key
DEEPSEEK_API_KEY=your_key
# Main LLM Configuration
LLM_PROVIDER=google
LLM_MODEL=gemini-2.5-flash-preview-05-20
# Motive Force (Autopilot) LLM Configuration (optional - defaults to main LLM)
LLM_PROVIDER_MOTIVE_FORCE=deepseek
LLM_MODEL_MOTIVE_FORCE=deepseek-chat
# Services
CHROMA_URL=http://localhost:8000
NEO4J_URI=bolt://localhost:7687
NEO4J_PASSWORD=password123
# Autopilot
MOTIVE_FORCE_ENABLED=false
MOTIVE_FORCE_MAX_CONSECUTIVE_TURNS=10
MOTIVE_FORCE_TEMPERATURE=0.8Enable via UI toggle. Your next message becomes the objective:
Using timestamps and normal querying, organize today's memories into 5-10 groups.
Delete redundant items, consolidate similar ones, add insights. Check with autopilot
periodically. Daily maintenance cultivates curated memory over time.
Configure via gear icon: turn delays, limits, memory integration, aggressiveness modes.
# Development
npm run dev # Full stack + KG sync
npm run dev:debug # With Node debugging
npm run dev:next # Frontend only
npm run dev:kg # KG sync only
# Knowledge Graph
npm run kg:sync # One-time sync
npm run kg:sync:full # Complete resync
npm run kg:sync:queue # Process retry queue
# Testing
npm run test # All tests
npm run test:rag # RAG system
npm run test:neo4j # Neo4j integrationskynet-agent/
├── src/
│ ├── app/ # Next.js routes
│ ├── components/ # React components
│ ├── lib/ # Core libraries
│ │ ├── motive-force-graph.ts # LangGraph workflow
│ │ ├── conscious-memory.ts # Volitional memory
│ │ ├── rag.ts # Automatic memory
│ │ └── knowledge-graph-*.ts # Neo4j integration
│ └── types/ # TypeScript definitions
├── docker-compose.yml # Services setup
└── motive-force-prompt.md # Autopilot personality
interface Memory {
id: string;
text: string;
embedding: number[]; // Google text-embedding-004
metadata: {
sender: 'user' | 'assistant';
timestamp: string;
summary?: string; // Auto-summarized if over limit
};
}interface ConsciousMemory {
id: string;
content: string;
tags: string[];
importance: number; // 1-10
source: 'explicit' | 'suggested' | 'derived';
metadata: {
accessCount: number;
lastAccessed: string;
};
}interface MotiveForceGraphState {
messages: BaseMessage[];
currentPurpose: string;
subgoals: SubGoal[];
executionPlan: ExecutionStep[];
toolResults: ToolResult[];
reflections: Reflection[];
overallProgress: number;
blockers: string[];
needsUserInput: boolean;
}POST /api/conscious-memory
{
"action": "save|search|update|delete|stats|tags",
"content": "string",
"tags": ["array"],
"importance": 7
}POST /api/motive-force
{
"action": "generate|generateStreaming|saveConfig|getState",
"sessionId": "string",
"data": {}
}- Semantic: Vector similarity via embeddings
- Keyword: Exact match fallback
- Smart Merge: Intelligent ranking with deduplication
- Automatic extraction from chat history
- Background service with retry queue
- Metrics collection and error handling
- Eventually consistent with ChromaDB
- Turn limits and error counting
- Manual override capabilities
- Resource usage monitoring
- Emergency stop functionality
"Embeddings service unavailable": Falls back to hash-based embeddings. Check Google API key.
"ChromaDB connection failed": Ensure docker-compose up -d and port 8000 available.
"Neo4j sync errors": Check credentials, run npm run kg:sync:queue for retries.
"Actually Looks Very Ugly": I suck at UI design.
Inspired by cognitive science:
- Dual-Process Theory: Automatic vs controlled processes
- Memory Consolidation: Active organization
- Working Memory: Conscious manipulation
Technical innovations:
- Hybrid Search: Solves subset query limitations
- MCP Architecture: Natural language memory control
- Importance Weighting: Smart prioritization
- LangGraph Integration: Complex autonomous workflows
Fork, improve, PR. Areas: memory algorithms, UI/UX, MCP tools, autopilot intelligence, testing, performance.
MIT - Lok Tar Ogar!
ChromaDB, Google AI, Anthropic MCP, Next.js, Neo4j teams. Open source MCP servers and Ollama Vercel AI SDK library.