Kept turns AI chats into local Markdown knowledge vaults

Kept is a new local-first knowledge manager for AI conversations. The project saves chats from ChatGPT, Claude, Gemini, Grok, and Kimi as Markdown files on the user’s own machine, then adds a desktop app for search, browsing, graph views, and reuse.

According to the project description, the archive lives under ~/.kept/ and includes an Obsidian-compatible vault. Kept also builds local indexes for full-text search, topics, projects, and graph views, so users can treat old chats as a structured knowledge base instead of leaving them trapped in vendor interfaces.

The tool is designed around the idea that AI chats often contain useful working memory: debugging trails, research notes, product decisions, prompts, snippets, and partial ideas. Kept turns those conversations into plain files the user owns, with the saved content organized by provider.

The project said it does not scrape rendered chat pages. Instead, a Chromium browser extension reads conversation data from provider API endpoints using the user’s signed-in session, normalizes the messages and supported image assets, and sends them to the Kept desktop app over http://localhost:18241.

On the desktop side, Kept writes Markdown files to ~/.kept/vault/, indexes them in SQLite using FTS5, and builds a local knowledge graph in CozoDB. The app also exposes search, graph, project, export, and agent tools.

Kept includes an MCP server called kept-vault-server. MCP, or Model Context Protocol, lets compatible agentic tools read and manage the vault, including listing, reading, writing, renaming, deleting, and searching files. The project said this works with clients such as Claude Code and OpenClaw.

The release supports Chrome, Brave, Edge, Vivaldi, Arc, Opera, and other Chromium-based browsers. The browser extension must be loaded manually in developer mode for now, while the desktop app is installed from GitHub Releases for macOS, Windows, or Linux.

On first launch, Kept creates the vault, local databases, a config file, and a token. The browser extension uses that token when sending data to the local server, which keeps the archive tied to the local installation.

The project said there is no hosted account and no cloud sync. Optional chat, topic discovery, and graph features use either the configured model provider or a local Ollama server, while the underlying archive remains stored as plain files on the user’s machine.

For users who want to inspect or edit the data directly, the Markdown files can be opened in Obsidian, VS Code, or any other editor. The project said the search and graph databases can be rebuilt from the files if needed, making the saved conversations the source of truth.