Memory Status

Memory Status is the read-only sibling in ComfyUI-MemoryCleaner. It doesn't clean anything, free anything, or trim anything - it just reads your RAM and VRAM and reports it, which makes it the more honest half of the pack. If you've ever had a "memory cleaner" node claim it freed 40GB and wondered how to check, this is the node you wire in first.

What it tells you

Run it and it prints a report to the ComfyUI console (the terminal window where you launched it) that looks like:

MEMORY STATUS
Process RAM: 12.34 GB (18.5%)
System RAM: 31.42 / 63.85 GB (49.2%)
VRAM Status:
  GPU 0 (NVIDIA GeForce RTX 4090):
    Allocated: 9.12 GB
    Reserved: 11.80 GB
    Total: 23.64 GB
    Usage: 38.6%

It uses psutil for the RAM side (your ComfyUI process's RSS plus whole-system numbers) and PyTorch's cuda.memory_allocated / cuda.memory_reserved for each GPU. The allocated-vs-reserved distinction is actually useful: reserved is what PyTorch has grabbed from the driver, allocated is what your models actually occupy, so the gap tells you how much headroom the CUDA caching allocator is sitting on.

Inputs and outputs

There's exactly one input, anything, and it's a pass-through - connect any data flow through the node and it flows out unchanged on output. That's how you drop a status check into the middle of a workflow without disturbing it. The output you care about is status_text, a STRING containing that whole report. Wire it into a text-display node (the built-in Show Text or whatever display node you have) and you get the report on the canvas instead of buried in the console. That's the move.

Why you'd actually use it

Three honest use cases:

• Debugging memory creep. Batch loops that slowly eat RAM are the classic ComfyUI failure mode. Drop a Memory Status at the end of each iteration and watch whether the numbers march upward.
• Verifying a cleaner actually cleaned. Put it before and after the MemoryCleaner node from this same pack and compare status_text reports. It's the difference between vibes and evidence.
• Checking headroom before a big generation. Wire it before a heavy VAE decode or video model to see whether you're about to OOM.

Install

Same pack as Memory Cleaner, so same install. ComfyUI Manager → search ComfyUI-MemoryCleaner, or:

cd ComfyUI/custom_nodes
git clone https://github.com/eddyhhlure1Eddy/ComfyUI-MemoryCleaner

then restart. Requirements are just psutil and torch - nothing heavy, no model files.

Gotchas

• If the console prints "VRAM: No CUDA devices available", it means torch doesn't see a CUDA device - you get the RAM half only. That's a torch/driver problem, not a node problem.
• The console print is the most complete version; if you only look at the canvas via status_text, remember it's a snapshot from the moment the node ran. Stale if you don't rerun.
• Since the pack is young and single-maintainer, the same advice as its sibling: it's a small, readable nodes.py - a two-minute skim settles any trust question.

Honestly, this is the node I'd reach for first out of the pack. It's free, it can't hurt anything, and it gives you a before/after baseline for every other memory trick you try - including the sledgehammer sitting next to it.