Knoverge

What it is

Knoverge is an open-source, self-hosted knowledge base that several AI agents and people share through MCP and HTTP APIs. It is not just long-term memory but a knowledge ledger: the canonical, auditable record of what a person, team or project knows. Every important change can be traced to the actor who made it and the source behind it. Changes are versioned and can be reviewed and reversed.

Why it exists

One person can work with many independent AI systems: ChatGPT, Claude Code, Codex, IDE agents, local models and custom automations. Each keeps its own context and memory. I built Knoverge as one shared knowledge layer for all of them, with people deciding what becomes canonical.

How it works

1. An agent discovers something and checks what the workspace already knows.
2. It proposes a new item or an update.
3. The server detects duplicates and conflicts.
4. A policy decides whether the change needs review.
5. A person or trusted curator approves it.
6. Canonical knowledge is updated, and every other agent can retrieve the new version.

Key decisions

Self-hosted first. One application container plus PostgreSQL is a complete installation, with no SaaS dependency.

MCP-native. An MCP endpoint at /mcp and the HTTP API expose the same twenty-nine tools from a single contract.

People stay in control. Agents can be read-only, propose-only or trusted. Nothing an agent writes becomes canonical without a rule that allows it, and the review queue shows why each proposal is waiting and what it would change.

Git as the history. Each item is a Markdown file in a real Git repository, with one commit per change. The repository can be read without Knoverge.

Provenance that cannot be rewritten. Every create, update, proposal and approval goes to an append-only, hash-chained event ledger.

Summaries are derived, never canonical. A summary records the revisions it was made from and goes stale when one of them changes.

Works with hosted assistants. Knoverge acts as its own OAuth 2.1 authorisation server, so ChatGPT and Claude.ai connectors can connect without a pasted token.

Current state

In development, at version 0.6.0 under the Apache-2.0 licence. The repository’s changelog and implementation plan list what each release contains. Still open: reading scanned PDFs, importing ChatGPT exports, knowledge “gardening”, and end-to-end browser tests.