Meko Documentation

The context engine for agentic applications

Shared, self-learning context for every agent, tool, and teammate: better answers, fewer tokens, one governed datapack.

Quick start

Connect your agent, add knowledge, watch Meko learn, and share with your team.

Get started →

Works with: WorkbenchLM (in-portal chat) | Claude Code | Cursor | Codex | Claude Desktop | Kiro | OpenCode | ...and more

What Meko is

Meko sits outside any single model or tool.

Memory built into Claude, ChatGPT, or Gemini stays inside that product and that agent's context. Meko is the data layer your agents share: knowledge, memory, and conversation history in one datapack that any agent, in any client, can read and write. Share datapacks across Claude, Cursor, Codex, custom agents, and the people on the team. You decide what is shared and with whom.

Agents reach it through one remote MCP endpoint. The same datapack holds memory, knowledge, conversations, and the trace of how an answer was produced.

Ask anything about Meko

Get instant answers from the docs, powered by AI.

Powered by kapa.ai

What you can do

Here are some of the things you can do with Meko:

Give every agent the same memory

Tell one agent a preference. A later session reads it back, including an agent in a different client, as long as both use the same datapack. See Watch your agent learn.

Connect the clients you already use

Cursor, Claude Code, Claude Desktop, Codex, Kiro, and OpenCode share one remote MCP endpoint. Each one reads and writes the same datapack. See Connect to an AI agent.

Search memory, knowledge, and past conversations in one call

context_search returns the pieces that match, so the agent does not have to load the whole store. On a partial match it uses what came back and says what is missing.

Give the agent documents to search

Add a file to the knowledge base. Later questions retrieve the passages that match.

Share a datapack with your team

Invite someone, or send a share link. They connect their own agents to the datapack you shared. A viewer reads it, a contributor adds memories, and a maintainer edits the knowledge base and manages members. See Share a datapack.

Promote a memory so the whole team gets it

A memory starts private. Promote it from Learnings and it becomes knowledge every member's agents can retrieve.

See whether a search found enough

The Self-learning tab shows which searches came back short. Add the missing fact once, then run the same search again.

Hand a file to the next agent

Artifacts store a file and return it whole, instead of pasting it back into the prompt.

See how an answer was produced

Traces record the reasoning, the tool calls, and the tokens behind a reply.

Try a feature in Labs

Labs walk a stage of the agentic development lifecycle in the portal, on a live demo application connected to a real datapack. Nothing to install.

Why use Meko

More context doesn't produce a better answer by itself. Too little and the model guesses. Too much and the right facts are buried while the token cost climbs.

Each agent and client also keeps its own partial context, so the next one relearns what another agent already knew.

The usual fix is a stitched stack: a database, a vector store, a RAG pipeline, a memory layer, and separate stores for conversations and artifacts. Each is tuned on its own, with no shared ranking and no single trace of where an answer came from.

A Meko datapack replaces that stack. One agent's learning is available to the others. Search ranks what comes back, so the model sees the useful pieces rather than the whole store. Every retrieval, memory update, and knowledge share is traced.

How Meko works

Meko datapack diagram

A datapack is the unit of isolation. It replaces the database, vector store, RAG pipeline, memory layer, and conversation store you would otherwise stitch together. It holds:

  • Memory — private facts, and the ones the team promotes into shared knowledge.
  • Knowledge — documents retrieved by hybrid search.
  • Decision traces — the reasoning, the tool calls, and the tokens behind an answer.
  • Artifacts — files an agent stores and reads back whole.

Every client uses the same MCP URL. A tool call names the datapack. Agents ask across memory, knowledge, and past conversations using context_search. Meko stores the turn, extracts durable facts into memory, and records the trace. The next session, on any connected client, searches that same datapack.

Underneath, Meko runs on YugabyteDB: vector, relational, and full-text search in one database. You use the MCP tools, not the database. See What is Meko to learn where it sits in an agent stack.

Learn more about Meko

Start where you fit