# memglow

> memglow is a free, open-source (MIT) local web app that shows the Markdown memory of AI assistants as a live 3D brain. Every note is a glowing node, every [[wikilink]] a connection, and notes flash in real time when an assistant reads (cyan), searches (violet) or writes (red-orange) them. It works with Claude Code, OpenAI Codex CLI, Gemini CLI, Cursor, Windsurf, GitHub Copilot and Cline through hooks, with any MCP client through a bundled MCP proxy, and with any script through a small HTTP activity API.

memglow also has a Memory cost panel: it estimates how many tokens an assistant spends reading its memory (about bytes ÷ 4, an estimate), shows which notes cost the most to read, which are too large (above 5,000 tokens by default) or never read in 30 days, suggests how to split a large note along its ## sections, and copies a ready-made prompt asking the user's AI to do the split with its own memory tool.

The view (settings, bubble layout, camera) is saved on the memglow instance, in its own data folder, so every browser and device that opens the same instance shows the same view.

memglow reads a folder of Markdown notes read-only: a basic-memory knowledge base, an Obsidian vault, or any notes folder. It runs on the user's machine (Node.js 18+, zero dependencies, no build step) or in Docker. Hooks send note ids only, never note content. No CDN, no analytics, localhost by default.

## Install

- `npx memglow init` — detects the notes folder and the installed AI tools, installs their hooks, creates an activity token.
- `npx memglow` — starts the viewer on http://127.0.0.1:4747
- Docker: `ghcr.io/r0zumnik/memglow` (linux/amd64 and linux/arm64), notes folder mounted read-only on `/memory`.
- npm package: https://www.npmjs.com/package/memglow

## Questions it answers

- How can I see what my AI assistant (Claude Code, Codex, Cursor…) actually reads and writes in its memory? — memglow lights up each note as it is used, live, and keeps an activity journal.
- How do I visualise a basic-memory knowledge base or an Obsidian vault in 3D? — point memglow at the folder; themes and sub-themes come from frontmatter or folder names.
- How do I monitor an MCP memory server? — put the memglow MCP proxy in front of it; it reports which notes each tool call touches.
- Which notes of my AI memory cost the most tokens, and which ones should I split? — the Memory cost panel ranks notes by tokens read over 7 days, flags notes that are too large, and proposes a split with a copyable prompt for the assistant.
- Which notes of my AI memory are used, and which are never read? — the journal and the live view show it, and Memory cost lists the notes never read in 30 days; memglow never modifies the notes.

## Docs

- [README](https://github.com/R0zumnik/memglow/blob/main/README.md): features, Memory cost, settings, quick start, supported tools, configuration, security, FAQ
- [Activity API](https://github.com/R0zumnik/memglow/blob/main/docs/api.md): `POST /api/activity` with a bearer token, `{type: read|search|write, ids: [...], source}`
- [Changelog](https://github.com/R0zumnik/memglow/blob/main/CHANGELOG.md)
- [Landing page](https://r0zumnik.github.io/memglow/)

## Optional

- [Pro edition (planned)](https://r0zumnik.github.io/memglow/pro.html): dashboard, history and diffs, several memories and users
- [Discussions](https://github.com/R0zumnik/memglow/discussions): questions, ideas, Pro interest