⚡ MNEURIX OPEN SOURCE UTILITY · WINDOWS PORTABLE EXE

MemScope

Live RAM / VRAM Pressure Monitor & Automated AI Watchdog for Windows.

Windows "Memory Compression" working set is a hidden pressure symptom — no stock tool surfaces it next to GPU VRAM and per-process hogs in one place. MemScope makes memory pressure visible, actionable, and self-healing for AI developers.

Download MemScope v1.1.0 (.EXE) ⬇ View Source on GitHub ↗

GPU & System Memory Watchdog

Zero-OOM Protection for Local LLM & Stable Diffusion Nodes

Real-time 64-bit DXGI VRAM tracking per process prevents out-of-memory crashes when hosting Ollama, llama.cpp, or ComfyUI workflows alongside background desktop applications.

MemScope v1.1.0 — Windows Memory Pressure Monitor

● PRESSURE: LIGHT (4.2 GB Compressed)

System RAM 55% Used

34.8 / 63.6 GB

Compressed Working Set: 10.0 GB

Dedicated GPU VRAM (24 GB) 49% Used

11.8 / 24.0 GB

Shared Memory: 0.68 GB GPU Util: 2.4%

PID	Process Name	RAM (RSS)	Private Working Set	VRAM Allocation	Watchdog Status
18492	`local-llm-engine.exe`	2.1 GB	24.2 GB	8.4 GB	● Active (Monitoring)
21048	`python-inference.exe`	4.2 GB	16.8 GB	3.2 GB	● Auto-Flush Enabled
28804	`dwm.exe` (Desktop Window Manager)	412 MB	512 MB	304 MB	System Core
14209	`browser.exe` (28 processes)	3.8 GB	4.6 GB	187 MB	User App

Why MemScope Exists — The Memory Compression Problem

When running local LLMs (Ollama, vLLM, llama-server) or AI image pipelines (ComfyUI, Stable Diffusion), GPU VRAM often spills over into system RAM. Windows responds by aggressively engaging Memory Compression.

The Hidden Bottleneck: Task Manager shows system RAM as "50% free", but hidden inside that allocation is 15 GB of compressed working sets. When your AI model accesses pages, CPU decompression stalls cause silent 10× generation slowdowns.
VRAM Truncation Bugs in Win32 API: Standard Windows tools relying on 32-bit WMI Win32_VideoController report a max of 4 GB VRAM on modern 16 GB and 24 GB graphics cards (such as high-VRAM AMD or NVIDIA GPUs). MemScope uses native 64-bit DXGI IDXGIAdapter::GetDesc() calls to surface true physical VRAM bounds.
Process-Level VRAM Allocation: MemScope maps Windows performance counters (\GPU Process Memory(*)\Local Usage) directly to active process IDs so you can instantly pinpoint which AI model container is hoarding VRAM.

Key Architectural Features

⚡ Real-Time Pressure Index

Derives a unified 0–100 Memory Pressure Index calculated from compressed MB, available system RAM, pagefile swap rates, and rate-of-change flags (Idle, Light, Moderate, High, Critical).

🤖 Automated AI Watchdog

Monitors local AI servers (`ComfyUI`, `llama-server`, `Ollama`). Automatically triggers memory working set flushes (`POST :8000/free`) or restarts deadlocked processes before OOM crashes occur.

🖥️ Mini HUD Overlay & Tray Icon

Includes a frameless, always-on-top draggable mini-overlay (~220×64px) and color-coded system tray icon so you can monitor VRAM while working in full-screen IDEs or 3D viewports.

📦 Single Portable Binary

Packaged as a standalone Windows executable (`MemScope-v1.1.0.exe`). Zero installation, zero administrative registry clutter, and zero background telemetry.

Feature Matrix Comparison

MemScope vs. Standard System Diagnostics
Feature	MemScope	Windows Task Manager	Process Hacker / System Informer	MSI Afterburner
Windows Memory Compression Counter	✓ Explicit	Hidden in sub-tab	Raw counter only	✗ Unsupported
64-bit DXGI VRAM Total (>4GB GPUs)	✓ True VRAM	✓ Supported	✓ Supported	✓ Supported
Per-Process VRAM Mapping (PID → GPU)	✓ Live Per-PID	Shared GPU column	✓ Supported	✗ Unsupported
Automated ComfyUI / LLM Watchdog	✓ Built-in	✗ Unsupported	✗ Unsupported	✗ Unsupported
Always-on-Top Mini HUD Overlay	✓ Draggable HUD	✗ Unsupported	✗ Unsupported	Requires RivaTuner

Download & Build Options

MemScope is published as free software under the Apache License 2.0 and hosted on GitHub (akketix/memscope) ↗. You can download the pre-compiled single binary or run directly from source using Python 3.11+ and uv.

Download MemScope v1.1.0 EXE (17.8 MB) ⬇ GitHub Repository & Source ↗

Open-Source Dependencies & Licenses

MemScope relies on open-source libraries that make fast, native Windows hardware diagnostics possible. All dependencies use permissive licenses:

Component / Dependency	Purpose in MemScope	Open-Source License
MemScope Core	Application logic, Watchdog, Pressure Index algorithm	Apache License 2.0
PySide6 (Qt6)	Desktop GUI controls, tray icon, and mini-overlay HUD	LGPL v3
psutil & pywin32	Process memory RSS, PIDs, and Windows PDH counters	BSD 3-Clause / PSF
pyqtgraph	Real-time time-series pressure history plotting	MIT License
FastAPI & uvicorn	Local REST API server and Prometheus metrics endpoint	MIT / BSD 3-Clause
LibreHardwareMonitorLib	Optional in-process hardware sensor reader (temp/fans)	Mozilla Public License 2.0 (MPL)

🎁 Open-Source & Free Community Utility