Konteks is a memory engine for AI coding agents.
It builds a project-local context graph through autonomous knowledge curation, ensuring you never re-explain your project to an AI agent.
Memory artifacts are stored directly inside your repository, exposing compact, task-specific recall through an MCP server without requiring global installation or cloud dependencies.
- Zero-Install: Run anywhere via
npxorbunxwithout global installation. - Language-Aware: Understands code structure across various file types.
- Local-First: Your project memory stays in your repo—no cloud, no accounts.
- Token-Efficient: High-fidelity context synthesis designed for LLM economy.
Konteks treats a repository as a local memory system for agent workflows. It builds searchable project artifacts from source files, retrieves task-relevant evidence, and expands related context through a lightweight graph of entities and relations when available.
For a deep dive into the philosophy, architecture, and usage, see the Full Documentation.
- Overview: Vision, Philosophy, and the "Why."
- Session Lifecycle: How to work with Konteks (Warm Up -> Build -> Save).
- Architecture Overview: How the memory engine works under the hood.
- Glossary: Short definitions for Konteks terms.
Konteks runs on Node.js (>=22) or Bun. Start by initializing memory from your project root:
npx -y konteks-cli initContinue with the Quickstart for MCP setup and the Warm Up -> Build -> Save flow.
If you are using an AI coding agent such as Claude Code, Codex, or OpenCode, let the agent handle the setup from your project root.
Paste this prompt into your agent. It points the agent to the Konteks installation playbook, asks it to follow project defaults, and keeps MCP setup, verification, and lifecycle guidance in one flow:
I want to add Konteks to this project.
Read and follow https://raw.githubusercontent.com/konteks/cli/refs/heads/main/docs/for-agents/installation.md. Treat that Markdown as the source of truth for installing Konteks, configuring MCP, verifying setup, and explaining the Warm Up -> Build -> Save workflow.
The agent should read the instructions, make safe defaults, ask only when a local-machine or agent-configuration decision is required, and leave you with a verified Konteks setup.
Konteks writes local memory under .konteks/. It uses SQLite for extracted content, durable memories, diary entries, retrieval indexes, embeddings, and graph metadata; extraction manifests and project config live beside the database.
MIT Licensed. See LICENSE for details.