Kyros

Persistent Memory System for AI Agents

Give your AI agents secure, self-correcting, persistent memory with cryptographic integrity and natural forgetting curves.

Features • Quick Start • Documentation • Examples • Contributing • License

---

Overview

Kyros is an open-source memory operating system designed specifically for AI agents. It provides three biologically-inspired memory types (episodic, semantic, and procedural) with built-in features like cryptographic integrity proofs, Ebbinghaus decay curves, and belief propagation.

The Problem

Modern AI agents are stateless — they forget everything between sessions. Developers manually implement memory using:

• Vector databases for semantic search
• Key-value stores for facts
• Custom logic for memory management
• Ad-hoc solutions for data integrity

This approach is fragile, insecure, and doesn't scale.

The Solution

Kyros provides a complete memory system in a single SDK:

from kyros import KyrosClient

client = KyrosClient(api_key="mk_live_default_dev_key_123456")

# Store a memory
client.remember("agent-1", "User prefers Python for backend development")

# Recall relevant memories
results = client.recall("agent-1", "What language does the user prefer?")
print(results.results[0].content)
# → "User prefers Python for backend development"

---

Features

Three Memory Types

Based on cognitive science research, Kyros implements three distinct memory systems:

• Episodic Memory: Events and experiences (conversations, actions, observations)
• Semantic Memory: Facts and knowledge (user preferences, domain knowledge)
• Procedural Memory: Skills and workflows (how to perform tasks)

Cryptographic Integrity

Every memory is protected with SHA-256 hashing and Merkle tree proofs:

• Detect tampering and poisoning attacks
• Verify memory authenticity
• Audit trail for all changes
• Immutable memory history

Ebbinghaus Decay

Memories naturally fade over time based on category-specific decay rates:

• Market data: 1.4 days half-life
• User identity: 693 days half-life
• Prevents memory bloat
• Keeps context relevant

Belief Propagation

When facts conflict, confidence updates ripple through the semantic graph:

• Automatic conflict resolution
• Confidence scoring
• Relationship tracking
• Self-correcting knowledge base

High Performance

• <20ms average recall latency
• <50ms memory storage
• pgvector for semantic search
• Redis caching layer
• Horizontal scaling support

Framework Integrations

Works seamlessly with popular AI frameworks:

• LangChain: Memory integration
• LlamaIndex: Data connector
• AutoGen: Multi-agent memory
• CrewAI: Crew memory
• Vercel AI SDK: Streaming support

Model Agnostic

Compatible with any LLM provider:

• OpenAI (GPT-4, GPT-3.5)
• Anthropic (Claude)
• Google (Gemini)
• Open source models (Llama, Mistral)
• Custom models

Built-in Model Context Protocol (MCP)

Kyros exposes a native Model Context Protocol (MCP) server:

• Connects directly to Agentic IDEs (Cursor, Windsurf, Cline) with a single command
• Exposes semantic memory recall, episode storage, and fact recording as local tools

Tenant, Project & Key Management

Designed for production environments, multi-tenant SaaS platforms, and enterprise control:

• Dynamic API Key Provisioning: Programmatically rotate, activate, and deactivate credentials
• Logical Namespaces: Isolate memories at the project, tenant, or agent levels dynamically
• Flexible Embedding Routing: Overwrite embedding models per request using the X-Embedding-Model header (with auto-padding for alignment with underlying database dimensions)

---

Quick Start

Prerequisites

• Docker and Docker Compose (for self-hosting)
• Python 3.11+ or Node.js 18+ (for SDK)

Installation

Option 1: Self-Hosted (Recommended)

# Clone the repository
git clone https://github.com/Kyros-494/kyros-ai.git
cd kyros-ai

# Start the server
docker compose up -d

# Server runs at http://localhost:8000

Option 2: Cloud Hosted (Sandbox)

Connect instantly to the managed sandbox by configuring the client with the sandbox flag:

from kyros import KyrosClient

client = KyrosClient(use_sandbox=True)

Option 3: Model Context Protocol (MCP) Server

For local Agentic IDEs (Cursor, Cline, Windsurf), run the built-in MCP server:

kyros mcp start

SDK Installation

Python:

pip install kyros-sdk

TypeScript:

npm install @kyros.494/sdk

Basic Usage

Python:

from kyros import KyrosClient

# Initialize client
client = KyrosClient(
    base_url="http://localhost:8000",
    api_key="mk_live_default_dev_key_123456"
)

# Store episodic memory (what happened)
client.store_episode(
    agent_id="agent-1",
    content="User asked about pricing plans",
    metadata={"timestamp": "2026-05-03T10:30:00Z"}
)

# Store semantic memory (what is true)
client.store_fact(
    agent_id="agent-1",
    subject="user_123",
    predicate="subscription_plan",
    value="Pro"
)

# Store procedural memory (how to do things)
client.store_procedure(
    agent_id="agent-1",
    name="handle_refund_request",
    steps=["Verify purchase", "Check refund policy", "Process refund"]
)

# Query memories
results = client.query(
    agent_id="agent-1",
    query="What plan is the user on?",
    memory_types=["semantic"]
)

for memory in results.results:
    print(f"{memory.content} (confidence: {memory.confidence})")

TypeScript:

import { KyrosClient } from '@kyros.494/sdk';

// Initialize client
const client = new KyrosClient({
  baseUrl: 'http://localhost:8000',
  apiKey: 'mk_live_default_dev_key_123456'
});

// Store and query memories
await client.storeEpisode({
  agentId: 'agent-1',
  content: 'User asked about pricing plans',
  metadata: { timestamp: '2026-05-03T10:30:00Z' }
});

const results = await client.query({
  agentId: 'agent-1',
  query: 'What plan is the user on?',
  memoryTypes: ['semantic']
});

results.results.forEach(memory => {
  console.log(`${memory.content} (confidence: ${memory.confidence})`);
});

---

Examples

LangChain Integration

from kyros.integrations.langchain import KyrosChatMemory
from langchain.chains import ConversationChain
from langchain.llms import OpenAI

# Initialize Kyros memory (api_key and base_url are resolved from environment variables if omitted)
memory = KyrosChatMemory(
    agent_id="chatbot-1"
)

# Create conversation chain
conversation = ConversationChain(
    llm=OpenAI(),
    memory=memory
)

# Chat with persistent memory
response = conversation.predict(input="My name is Alice")
# Memory automatically stored

# response = conversation.predict(input="What's my name?")
# → "Your name is Alice"

Multi-Agent System

from kyros import KyrosClient

client = KyrosClient(api_key="mk_live_default_dev_key_123456")

# Agent 1: Research agent
client.remember("researcher", "Found 3 relevant papers on neural networks")

# Agent 2: Writer agent queries researcher's findings
results = client.recall("researcher", "What papers did you find?")
papers = results.results[0].content

# Writer uses the information
client.remember("writer", f"Writing article based on: {papers}")

Causal Reasoning

# Store causal relationships
client.store_causal_edge(
    agent_id="agent-1",
    cause_memory_id="mem_abc123",
    effect_memory_id="mem_def456",
    confidence=0.95
)

# Query causal chains
chain = client.get_causal_chain(
    agent_id="agent-1",
    memory_id="mem_def456"
)

# Understand why something happened
for link in chain:
    print(f"{link.cause} → {link.effect} (confidence: {link.confidence})")

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                        Your Application                      │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                      Kyros SDK (Python/TS)                   │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    Kyros API Server (FastAPI)                │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐      │
│  │   Episodic   │  │   Semantic   │  │  Procedural  │      │
│  │    Memory    │  │    Memory    │  │    Memory    │      │
│  └──────────────┘  └──────────────┘  └──────────────┘      │
│                                                               │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐      │
│  │   Integrity  │  │    Decay     │  │    Belief    │      │
│  │    Proofs    │  │    Engine    │  │ Propagation  │      │
│  └──────────────┘  └──────────────┘  └──────────────┘      │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│              Storage Layer (PostgreSQL + Redis)              │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐      │
│  │  PostgreSQL  │  │    Redis     │  │   pgvector   │      │
│  │  (Memories)  │  │   (Cache)    │  │ (Embeddings) │      │
│  └──────────────┘  └──────────────┘  └──────────────┘      │
└─────────────────────────────────────────────────────────────┘

---

Documentation

Getting Started

• Introduction - Learn about Kyros and its architecture
• Quick Start Guide - Get up and running in 5 minutes
• Core Concepts - Understand memory types and features
• SDK Publication & LLM Integrations - Guide to packaging, local installations, and LLM integrations

SDK References

• Python SDK - Complete Python API reference
• TypeScript SDK - Complete TypeScript API reference

Deployment

• Self-Hosting Guide - Deploy Kyros on your infrastructure
• Configuration - Environment variables and settings
• Security Best Practices - Secure your deployment

Advanced Topics

• Architecture Deep Dive - System design and internals

---

Community

Get Help

• Documentation: docs.kyros.ai
• GitHub Discussions: Ask questions
• GitHub Issues: Report bugs
• Email: [email protected]

Stay Updated

• Star this repo to follow development
• Watch releases for new versions

Contributing

We welcome contributions from the community! See our Contributing Guide for:

• Code of Conduct
• Development setup
• Commit conventions
• Pull request process
• Testing requirements

Good First Issues: View beginner-friendly tasks →

---

License

Kyros uses an Open Core licensing model:

• Server Core: Apache 2.0 License - Free to self-host and modify
• SDKs & Integrations: MIT License - Free to use in any project
• Enterprise Modules: Commercial License - Contact us

This means you can:

• [Allowed] Use Kyros for free in commercial projects
• [Allowed] Self-host on your own infrastructure
• [Allowed] Modify the source code
• [Allowed] Distribute your modifications
• [Prohibited] Use the "Kyros" trademark without permission

See LICENSE and LICENSE-MIT for full terms.

---

Security

We take security seriously. If you discover a security vulnerability, please email [email protected] instead of using the issue tracker.

• Response Time: Within 48 hours
• Patch Timeline: 7 days for critical issues
• Disclosure: Coordinated disclosure policy

See SECURITY.md for our full security policy.

---

Acknowledgments

Kyros is built on the shoulders of giants:

• FastAPI - Modern Python web framework
• PostgreSQL - Reliable database
• pgvector - Vector similarity search
• Redis - High-performance caching
• Sentence Transformers - Embedding models
• LangChain - LLM application framework

Special thanks to all our contributors who make this project possible.

---

Built by the Kyros team

Website • Documentation • GitHub •

_{Copyright © 2026 Kyros. All rights reserved.}