Core concepts

Understanding these four concepts will make everything else in the docs click.

Memory

A memory is a discrete, structured piece of context derived from your AI conversations. It is not a raw transcript. It is a classified, ranked unit of information.

{
  content: "Prefers TypeScript. Dislikes class components.",
  type: "preference",
  domain: "engineering",
  importance: 0.8,
  created_at: "2026-03-27T00:00:00Z"
}

Memories are stored in your vault and retrieved selectively. Only what is relevant to the current session gets injected into the prompt. Not everything and not nothing.

The vault

Your vault is the storage layer where memories live. Today, with kontxt, that vault is a local SQLite database at ~/.kontxt/vault.db. The encrypted, permissioned vault (and hosted sync) are on the 4StaX roadmap. Consent logging, granular permissions, and revocation are also part of the Vault roadmap.

Relevance ranking

When your AI client needs context, kontxt doesn’t dump your entire memory archive into the prompt. It ranks candidates using four signals:

Signal	What it measures
Semantic similarity	How closely the memory matches the current task
Recency	How recently the memory was created or accessed
Access frequency	How often this memory has been useful before
Explicit importance	The importance score set when the memory was stored

The result is a compact, relevant bundle. It is not noise.

Memory types

Type	Examples
`fact`	Your name, role, team, tech stack
`preference`	Language choices, formatting preferences, communication style
`project`	Project names, goals, ongoing context
`decision`	Architectural decisions, tradeoffs, “we chose X because Y”
`skill`	Things you can do (and how you do them)
`episodic`	Session-specific context worth retaining for later

MCP and how it connects

MCP (Model Context Protocol) is the standard that lets AI clients talk to external tools and data sources. kontxt implements an MCP server that exposes your vault to any compatible client. When you ask Cursor or Claude Desktop something, the MCP client calls kontxt tools, retrieves relevant memories, and injects them into your prompt automatically. You do not see it happening. It just works.

Open source vs cloud

	kontxt (OSS)	4StaX cloud
Storage	Local SQLite	Hosted, encrypted
Sync	Single machine	Cross-device
Access	No account needed	Account required
Status	Available now	Waitlist

The MCP tool surface is identical between both. Integrations built on kontxt will work with 4StaX cloud without changes.

Get started

Memory

Vault

Developers

Core concepts

Memory

The vault

Relevance ranking

Memory types

MCP and how it connects

Open source vs cloud

​Memory

​The vault

​Relevance ranking

​Memory types

​MCP and how it connects

​Open source vs cloud

Memory

The vault

Relevance ranking

Memory types

MCP and how it connects

Open source vs cloud