How Meko fits in your stack

Where Meko sits in your architecture and how it deploys

Meko is an agent-native data layer for multi-agent systems that work and learn together. It supports:

  • Collective memory. Learning compounds across the entire system, not just per-agent memory.
  • Shared knowledge. Meko builds knowledge over time from conversations, real-time data sources, and slower-changing knowledge bases.
  • Auditability of the learning process. Connect execution traces to how agents process data and knowledge, learn from it, and share that learning.

Meko is serverless, multi-tenant (multi-agentic), and optimized for tiering to object stores. In addition to providing memory, conversation history, shareable knowledge, and full chain-of-thought traceability, Meko also exposes PostgreSQL-compatible database models (vector, SQL, NoSQL, graph, and more).

Architecture

Meko is fully open-source and integrates easily with any agentic framework through MCP servers. It is built on top of a unified distributed PostgreSQL data layer that supports vector, SQL, NoSQL, graph, and search.

Meko sits between your agent SDK and the underlying database, providing a unified data layer that replaces multiple standalone systems.

Meko Architecture Diagram

Agents interact via MCP. Developers interact via CLI, API, or direct PostgreSQL connection. Meko runs anywhere - local, cloud, on-premises, or air-gapped - with the same configuration everywhere.

What you get

Each agent gets an isolated datapack with:

Capability Description
PostgreSQL (YugabyteDB YSQL) Structured data and conversation history
Vector search (pgvector, hnswlib, scaNN) Embeddings and similarity search
Graph memory (mem0) Long-term entity and relationship extraction from conversations
RAG pipeline
(pg_dist_rag + unstructured + LangChain)		 Document chunking, embedding, indexing with built-in data preprocessing
Inference caching (LMCache / KDN) Reduces redundant LLM calls without a separate cache layer
MCP server Agents discover and query their own datapack
Audit trails LLM reasoning traces and execution logs
Observability, security, RBAC Spans the full stack, not bolted on per-component

The inference, knowledge, and database layers share state, and they operate as one system, not three isolated tiers.

What meko replaces

Without Meko, building an agentic application typically requires wiring together:

  • A relational database for structured data
  • A vector database for embeddings and similarity search
  • A graph database for entity relationships
  • An object store for conversations
  • A separate cache layer for inference
  • Custom integration code for each system
  • Per-component observability, security, and RBAC

Meko replaces all of this with a single system. The inference, knowledge, and database layers share state and operate as one.

How agents connect

Agents interact with Meko via MCP (Model Context Protocol). Each datapack exposes an MCP server that provides tools for memory, knowledge, database access, and more.

Developers can also interact via:

  • CLI: meko commands for datapack management, memory operations, and knowledge base administration.
  • REST API: Programmatic access to all datapack operations.
  • Direct PostgreSQL: Connect to your datapack's YugabyteDB YSQL database with standard PostgreSQL tools.

Framework compatibility

Meko works with any agent framework that supports MCP:

  • CrewAI
  • LangChain / LangGraph
  • OpenAI Agents SDK
  • Any MCP-compatible client

Deployment

Meko runs anywhere:

  • Local development: meko start to run everything on your machine.
  • Cloud: Deploy as a managed service or self-hosted on any cloud provider.
  • On-premises / air-gapped: Same system, no external dependencies required.

The same configuration works everywhere, with no architectural rewrites between development and production.

Scale out by adding YugabyteDB nodes, and scale down by removing them, with no downtime.

Built on

Meko is built on proven open-source technologies:

Component Technology
Distributed database YugabyteDB
Vector search pgvector
Graph memory mem0
RAG pipeline pg_dist_rag + unstructured + LangChain
Observability LangFuse
Inference serving vLLM
Inference caching LMCache