Overview - AG-Kit Documentation

AG-Kit provides a sophisticated dual-layer memory management system that enables intelligent context management and knowledge storage for AI agents through coordinated short-term and long-term memory operations.

Memory System Architecture

AG-Kit’s memory system consists of two complementary layers that work together to provide comprehensive memory capabilities:

Short-Term Memory (Session Memory)

Stores conversation events and messages with efficient storage, retrieval, and search capabilities. Functions like traditional storage systems with precise querying and filtering based on various conditions. Key Characteristics:

Volatile or Persistent - Choose between in-memory or cloud storage
Session-based - Isolated contexts for different conversations
Token-aware - Automatic management of LLM context windows
Context Engineering - Built-in compaction and summarization for long conversations (Learn more)
Fast Access - Optimized for recent conversation history

Use Cases:

Maintaining conversation context within a session
Managing multi-turn dialogues
Tracking tool call history
Handling concurrent user conversations

Long-Term Memory (Knowledge Memory)

Intelligently extracts important information from conversations to store user profiles, key facts, and preferences. Supports semantic similarity-based recall using vectorization technology for efficient similarity search. Key Characteristics:

Persistent - Information survives across sessions
Semantic Search - Vector-based similarity matching
Intelligent Extraction - Automatic memory extraction from conversations
Knowledge Consolidation - Deduplication and merging of related information

Use Cases:

Remembering user preferences and facts
Building user profiles over time
Storing domain knowledge
Personalizing agent responses

Context Engineering

AG-Kit includes built-in context engineering to solve the context degradation problem in long-running conversations. As conversations grow, they face challenges like token limits, attention dilution, and performance degradation.

Three-Tier Management Strategy

Normal Operation (0-80% of threshold): Store all messages in full detail
Reversible Compaction (80-95% of threshold): Compress old messages while preserving reconstruction ability
Structured Summarization (95%+ of threshold): Create structured summaries for dramatic token reduction

Key Benefits

Unlimited Conversation Length: No practical limits on conversation duration
Maintained Quality: Performance remains high even in long conversations
Information Preservation: Critical context preserved through intelligent management
Cost Efficiency: Dramatic reduction in token usage and associated costs

Context engineering happens automatically - no manual intervention required. The system monitors token usage and applies the appropriate strategy based on configurable thresholds.

Memory Workflow

Process Flow

Short-Term Retrieval: Fetch relevant context from conversation history
Long-Term Recall: Semantic search for user profiles and important facts
Context Integration: Combine both memory layers for comprehensive context
Response Generation: LLM generates response with full context
Smart Extraction: Automatically identify and store important information
Continuous Learning: Continuously accumulate and optimize user knowledge

Memory Implementations

Short-Term Memory Implementations

InMemoryMemory

In-memory storage for development and testing

Fast read/write operations
Multi-session support
Content-based search
Zero external dependencies

TDAIMemory

Cloud-based storage for production

Persistent storage
Distributed sessions
Advanced search
Enterprise reliability

CloudBaseMemory

Tencent CloudBase cloud database

Serverless NoSQL storage
Real-time synchronization
Built-in authentication
Auto-scaling capabilities

MongoDBMemory

MongoDB document database

Flexible document storage
Rich query capabilities
Horizontal scaling
ACID transactions

MySQLMemory

MySQL relational database

ACID compliance
Mature ecosystem
High performance
Enterprise features

TypeORMMemory

TypeORM multi-database support

Database agnostic
Type-safe queries
Migration support
Multiple DB backends

Long-Term Memory Implementations

Mem0LongTermMemory

Mem0 SDK integration with AI-powered features

Automatic extraction
Semantic search
Graph relationships
Intelligent consolidation

TDAILongTermMemory

TDAI cloud storage for enterprise

Cloud persistence
Strategy-based organization
Scalable infrastructure
Enterprise features

Memory Comparison

Short-Term vs Long-Term Memory

Aspect	Short-Term Memory	Long-Term Memory
Purpose	Conversation history	Persistent knowledge
Scope	Single session	Cross-session
Storage	Recent messages	Extracted facts
Retrieval	Chronological/Search	Semantic search
Lifespan	Session duration	Indefinite
Size Limit	Token-based	Unlimited
Data Type	Raw messages	Structured facts
Update Frequency	Every message	On extraction
Primary Use	Context window	Personalization

Implementation Comparison

Feature	InMemory	TDAI	CloudBase	MongoDB	MySQL	TypeORM
Persistence	❌	✅	✅	✅	✅	✅
Setup	None	Medium	Medium	Low	Low	Medium
Performance	Excellent	Good	Good	Good	Excellent	Good
Best For	Dev/Test	Enterprise	Serverless	Documents	Traditional	Multi-DB

Storage Backend Guide

Quick Overview

Backend	Best For	Setup	Persistence
InMemory	Development, Testing	Zero config	❌
TDAI	Enterprise, Production	API key	✅
CloudBase	Serverless, Tencent Cloud	Cloud config	✅
MongoDB	Document-heavy, Flexible schema	Database setup	✅
MySQL	Traditional apps, ACID compliance	Database setup	✅
TypeORM	Multi-database, Type safety	ORM config	✅

Choosing the Right Backend

By Use Case

Development/Testing: InMemoryMemory
Enterprise Production: TDAIMemory, MySQLMemory
Serverless Apps: CloudBaseMemory, TDAIMemory
Document Storage: MongoDBMemory, CloudBaseMemory
Multi-Database: TypeORMMemory

Quick Start Examples

// Development
const memory = new InMemoryMemory({ sessionId: 'dev' });

// Production Cloud
const memory = new TDAIMemory({
  sessionId: 'user-123',
  clientOptions: { apiKey: process.env.TDAI_API_KEY }
});

// Self-hosted MySQL
const memory = MySQLMemory.create({
  host: 'localhost',
  database: 'agkit',
  sessionId: 'user-123'
});

Core Concepts

Session Management

Sessions provide isolated conversation contexts for different users or conversation threads.

// Create isolated sessions for different users
const aliceSession = 'user-alice-session-1';
const bobSession = 'user-bob-session-1';

// Each session has independent memory
await memory.add({sessionId: aliceSession, message, state});
await memory.add({sessionId: bobSession, message, state});

Learn more: Session Management Guide

Token Management

Automatic token counting and trimming to manage LLM context windows. For advanced strategies, see the Context Engineering Guide.

// Retrieve events within token limit
const events = await memory.list({
  maxTokens: 4000,  // Fits within GPT-4 context
  sessionId: 'session-123'
});

Learn more: Short-term Memory Guide

Memory Strategies

Organize long-term memories by strategy for better categorization.

// Different memory strategies
await longTermMemory.record({
  strategy: 'facts',        // Objective information
  content: 'User graduated from MIT'
});

await longTermMemory.record({
  strategy: 'preferences',  // User likes/dislikes
  content: 'User prefers dark mode'
});

Learn more: Long-term Memory Guide

Semantic Search

Find relevant memories using natural language queries.

// Search across all memories
const results = await longTermMemory.semanticSearch(
  'What are the user\'s food preferences?'
);
// Returns: "User loves Italian food", "User is allergic to shellfish", etc.

Learn more: Long-term Memory Guide

Getting Started

Choose Storage Backend

Pick based on your needs: InMemory (dev), TDAI/CloudBase (cloud), or MySQL/MongoDB (self-hosted).

Initialize Memory Layers

Set up short-term memory for conversations and/or long-term memory for knowledge.

Integrate with Agent

Attach memory instances to your agent configuration.

Implement Session Management

Create session IDs and manage session lifecycle.

Test and Optimize

Monitor performance and adjust token limits, caching, and cleanup strategies.

Next Steps

Short-term Memory

Conversation history and session management

Long-term Memory

Persistent memory with semantic search

Agent Integration

Complete guide to memory-agent integration

Session Management

Multi-user sessions and context isolation

API Reference

Complete API documentation and reference

Getting Started

Building Agents

Deploying Agents

Building Agent Apps

Open Protocols

Community

​Memory System Architecture

​Short-Term Memory (Session Memory)

​Long-Term Memory (Knowledge Memory)

​Context Engineering

​Three-Tier Management Strategy

​Key Benefits

​Memory Workflow

​Process Flow

​Memory Implementations

​Short-Term Memory Implementations

InMemoryMemory

TDAIMemory

CloudBaseMemory

MongoDBMemory

MySQLMemory

TypeORMMemory

​Long-Term Memory Implementations

Mem0LongTermMemory

TDAILongTermMemory

​Memory Comparison

​Short-Term vs Long-Term Memory

​Implementation Comparison

​Storage Backend Guide

​Quick Overview

​Choosing the Right Backend

​By Use Case

​Quick Start Examples

​Core Concepts

​Session Management

​Token Management

​Memory Strategies

​Semantic Search

​Getting Started

​Next Steps

Short-term Memory

Long-term Memory

Agent Integration

Session Management

API Reference

Memory System Architecture

Short-Term Memory (Session Memory)

Long-Term Memory (Knowledge Memory)

Context Engineering

Three-Tier Management Strategy

Key Benefits

Memory Workflow

Process Flow

Memory Implementations

Short-Term Memory Implementations

Long-Term Memory Implementations

Memory Comparison

Short-Term vs Long-Term Memory

Implementation Comparison

Storage Backend Guide

Quick Overview

Choosing the Right Backend

By Use Case

Quick Start Examples

Core Concepts

Session Management

Token Management

Memory Strategies

Semantic Search

Getting Started

Next Steps