DEV Community

Cover image for I built a live 3D view of my AI assistant's memory — and found where its tokens go
R0zumnik.RuFr
R0zumnik.RuFr

Posted on Originally published at github.com

I built a live 3D view of my AI assistant's memory — and found where its tokens go

#ai #opensource #productivity #claude

My AI coding assistants keep their long-term memory as plain Markdown notes (basic-memory, but an Obsidian vault works the same way). It works well — until you wonder what the assistant actually does with it. Which notes does it open? What does it search for? What does it overwrite?

I couldn't answer, so I built memglow: a small local web app that turns the memory folder into a live 3D graph. Every note is a node, every [[wikilink]] an edge, and notes light up as the assistant works — cyan when read, violet when searched, red-orange when written, with comets running along the links.

memglow demo

How the activity gets in

  • Hooks for Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Copilot and Cline — npx memglow init detects which ones you use and installs them.
  • An MCP proxy for any other client: it sits in front of your memory MCP server, relays everything unchanged and reports which notes were touched.
  • A tiny HTTP API for anything else.

Hooks send note ids only, never content. memglow runs on localhost, only reads your notes, and has zero runtime dependencies.

The surprise: where the tokens go

Once I could see the reading, a pattern jumped out: a handful of notes were read over and over, and some of them had quietly grown huge. So memglow 0.3 added a Memory cost panel:

  • tokens read today / over 7 days (≈ bytes ÷ 4 — an estimate, not a bill),
  • the notes that cost the most to read, and how many notes make up most of the reading,
  • oversized notes with their sections and a suggested split,
  • a "Copy prompt for your AI" button, so your own assistant does the split (memglow never writes your notes).

Memory cost

On my own memory, the note tracking this very project had grown to ~15,500 tokens in two days and was read in full every time. After the split, a typical read costs ~1,200 tokens for the summary, or 4–6k with one topic note. Your numbers will differ — the point is that the tool measures them on your memory.

Try it 

npx memglow init
Enter fullscreen mode Exit fullscreen mode

or with Docker: ghcr.io/r0zumnik/memglow. MIT licensed: https://github.com/R0zumnik/memglow

I'd love feedback — especially on what you'd want to see about your assistant's memory.

Top comments (0)