Overview

kumbuka is an open-source team memory system for AI assistants. It gives a team a durable, shared place for the knowledge an assistant should carry across conversations, and serves that knowledge to MCP-capable AI clients over a remote server. A web admin console lets the team curate it.

This page orients you. For the precise domain model see Concepts; for how an assistant calls it see the MCP tools.

What it is for

Teams using AI assistants re-explain the same steering knowledge in every session — the decisions, conventions, and constraints that should shape how the assistant works. kumbuka turns that into a first-class, team-owned asset so it is applied without being re-told.

It captures work-steering knowledge, deliberately small and typed:

• the rules and definitions that guide how the assistant works, not a copy of the team’s documents or source (those stay in their own systems);
• shared and curatable, so the team sees and edits what the assistant relies on rather than each person accumulating an opaque, divergent context;
• portable, so any MCP-capable assistant reads and writes it through one endpoint.

The personal / shared boundary

kumbuka has two halves, and the line between them is the product’s defining principle.

• Shared memory is what the team curates together — the global baseline and any number of project scopes. It is visible in the console and editable by the team according to role.
• Private memory is each member’s own working space. It is reachable only by that member, only through their own authenticated MCP session. No admin, no console screen, and no team-facing API can read it. This is enforced structurally, not by a setting (see Security & privacy).

When the two ever appear to conflict, the private guarantee wins over convenience.

Where it runs

kumbuka is a single Docker Compose stack: a Quarkus backend, PostgreSQL, Keycloak (the identity provider), and a Caddy edge, with the Next.js admin console. The deployable stack lives in kumbuka-server. See the Quickstart to self-host it and the Architecture for the topology.

Editions

The Community Edition documented here is the free, self-hosted, single-tenant memory core. A commercial path (multi-tenancy, the Context Documents extension, a moderation add-on, and hosted SaaS) is planned but pre-beta — see Editions. It is described honestly: no overclaiming, no prices, no dates.