Research
AI Agent Platform with persistent memory, self-learning capabilities, and workflow automation.
Persistent Memory
Agents remember user preferences across sessions
Self-Learning
Automatic knowledge acquisition from conversations
RAG System
Semantic search over documents using embeddings
System Architecture
The platform follows a layered architecture with clear separation of concerns.
┌─────────────────────────────────────────────────────────────────┐
│ PRESENTATION LAYER │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Dashboard │ │ Chat UI │ │ Workflow │ │
│ │ (Admin) │ │ (Widget) │ │ Builder │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ API LAYER │
│ /api/chat /api/agents /api/brain /api/workflows │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ SERVICE LAYER │
│ ChatService MemoryService KnowledgeService RAGService │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ REPOSITORY LAYER │
│ AgentRepo ConversationRepo MemoryRepo KnowledgeRepo │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ DATA LAYER │
│ MySQL Database Cloud Storage │
└─────────────────────────────────────────────────────────────────┘
Tech Stack
| Component | Technology | Rationale |
|---|---|---|
| Runtime | Node.js | Non-blocking I/O for real-time chat |
| Framework | Express.js | Lightweight, flexible routing |
| Database | MySQL | ACID compliance, JSON support |
| AI Model | GPT-4.1-mini | Balance of cost and capability |
| Embeddings | text-embedding-3-small | 1536 dimensions, cost-effective |
Memory System
The three-tier memory architecture enables agents to maintain context and learn from interactions.
┌─────────────────────────────────────────────────────────────────────────┐ │ AGENT INTELLIGENCE │ ├─────────────────────────────────────────────────────────────────────────┤ │ │ │ ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ │ │ SHORT-TERM │ │ LONG-TERM │ │ KNOWLEDGE │ │ │ │ MEMORY │ │ MEMORY │ │ BASE │ │ │ │ │ │ │ │ │ │ │ │ Conversation │ │ User Profile │ │ Q&A Pairs │ │ │ │ History │ │ Preferences │ │ Documents │ │ │ │ (Last N msgs) │ │ Facts │ │ (RAG) │ │ │ │ │ │ │ │ │ │ │ │ Scope: Session │ │ Scope: User │ │ Scope: Agent │ │ │ │ TTL: ~30 min │ │ TTL: Permanent │ │ TTL: Permanent │ │ │ └────────┬────────┘ └────────┬────────┘ └────────┬────────┘ │ │ │ │ │ │ │ └─────────────────────┼─────────────────────┘ │ │ ▼ │ │ ┌────────────────────────┐ │ │ │ CONTEXT BUILDER │ │ │ └────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌────────────────────────┐ │ │ │ OPENAI GPT │ │ │ └────────────────────────┘ │ └─────────────────────────────────────────────────────────────────────────┘
1. Short-Term Memory
Maintains conversation continuity within a session by storing the last N messages.
How It Works
- User sends message
- Message saved to
messagestable with conversation_id - On next request, last N messages retrieved
- History included in OpenAI API call
2. Long-Term Memory
Remembers user-specific information across sessions using a key-value store.
User: "Remember that I prefer morning appointments"
│
▼
┌────────────────────────────────────────────────────────────────┐
│ AI RESPONSE │
│ Detects memory request and outputs: │
│ {"action":"save", "key":"appointment_preference", │
│ "value":"prefers morning appointments"} │
└───────────────────────────┬────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ MEMORY REPOSITORY │
│ INSERT INTO memories (agent_id, session_id, key, value) │
└────────────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────────────┐
│ FUTURE CONVERSATIONS │
│ Context includes: │
│ - appointment_preference: prefers morning appointments │
└────────────────────────────────────────────────────────────────┘
3. Knowledge Base (Self-Learning)
Builds institutional knowledge from conversations through a training loop.
Step 1: User Asks Unknown Question
───────────────────────────────────
User: "What are your business hours?"
│
▼
┌───────────────┐
│ Agent doesn't │────► Log to unanswered_questions
│ know answer │
└───────────────┘
Step 2: Owner Trains Agent
──────────────────────────
┌─────────────────────────────┐
│ TRAINING DASHBOARD │
│ │
│ Pending Questions: │
│ "What are your hours?" │
│ Asked: 15 times │
│ │
│ Owner provides answer: │
│ "Mon-Fri, 9 AM to 5 PM" │
└──────────────┬──────────────┘
▼
┌─────────────────────────────┐
│ Add to knowledge base │
└─────────────────────────────┘
Step 3: Agent Uses Knowledge
────────────────────────────
Future questions get accurate answers automatically
RAG Pipeline
Retrieval Augmented Generation enables semantic search over uploaded documents.
INGESTION PHASE
───────────────
Document Upload Chunking Embedding
│ │ │
▼ ▼ ▼
┌─────────┐ ┌─────────────────┐ ┌─────────────────┐
│ PDF │ │ Split into │ │ OpenAI API │
│ CSV │ ────────► │ 800-char │ ───────► │ Embeddings │
│ TXT │ │ chunks │ │ (1536 dims) │
└─────────┘ └─────────────────┘ └────────┬────────┘
▼
┌───────────────────┐
│ Store in MySQL │
└───────────────────┘
RETRIEVAL PHASE
────────────────
User Query Query Embedding Similarity Search
│ │ │
▼ ▼ ▼
┌─────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ "What is │ │ Embed query │ │ Cosine │
│ the refund │ ────► │ using same │ ─────► │ similarity │
│ policy?" │ │ model │ │ search │
└─────────────┘ └─────────────────┘ └────────┬────────┘
▼
┌─────────────────┐
│ Inject into │
│ AI context │
└─────────────────┘
Document Chunking
Documents are split into overlapping chunks to preserve context across boundaries.
800
Chunk Size (chars)
200
Overlap (chars)
1536
Embedding Dims
Vector Embeddings
Text is converted to 1536-dimensional vectors that capture semantic meaning.
Input Text:
"We offer a 30-day money-back guarantee on all products"
│
▼
┌──────────────────┐
│ OpenAI API │
│ text-embedding- │
│ 3-small │
└────────┬─────────┘
│
▼
Output Vector (1536 dimensions):
[0.0234, -0.0891, 0.0123, 0.0567, -0.0345, ..., 0.0456]
Similar texts will have similar vectors (high cosine similarity)
Cosine Similarity Search
Queries are matched against stored chunks using cosine similarity.
Query: "What is your refund policy?"
- Chunk about refunds and returns 0.92
- Chunk about shipping policies 0.45
- Chunk about contact information 0.38
API Reference
Chat Endpoints
POST
/api/chat/:agentSlug/message
Send a message to an agent
{
"session_id": "user_abc123",
"message": "What are your business hours?"
}
Agent Management
GET
/api/agents
List all agents
POST
/api/agents
Create a new agent
PUT
/api/agents/:id
Update an agent
Workflows
POST
/api/workflows/ai-build
Build workflows from natural language
{
"message": "When someone submits a form, send me an email"
}