16 GB of RAM, nothing heavy running, yet Task Manager shows 90% memory used and everything crawls. A reboot helps for a few hours, then the creep starts again. If the growth tracks your use of Studio Effects, Recall, or live captions, you may be looking at an NPU driver memory leak — the driver allocates memory for AI work and never gives it back.
Memory leaks masquerade as ordinary slowness. Users blame Windows, buy more RAM, or reinstall everything — when the real culprit is a driver that loses a few megabytes per AI inference and never cleans up. Over a day of uptime, those megabytes become gigabytes.
This guide shows how to prove the leak is real, then fix it with six methods in technician order: confirm the diagnosis, update or replace the driver, and tame the workloads that trigger the worst leaks. No registry hacks, no snake oil — just the driver stack, done properly.
Why this happens
- The driver never frees inference buffers: Each AI workload allocates working memory on the host side. A buggy driver forgets to release some buffers when the workload ends, so every Studio Effects session permanently costs RAM.
- Handle leaks alongside memory leaks: Leaking thread handles, events, or file mappings alongside memory is the classic signature of a driver that skips cleanup — often visible in Task Manager before RAM numbers get scary.
- Leak triggered by specific features: Some leaks only occur with particular workloads — creative filters or long Recall sessions — making the leak seem intermittent.
- Driver/firmware mismatch: The driver expects the firmware to acknowledge buffer releases; mismatched firmware never does, so the driver holds the memory forever.
- Windows Update installed a leak-prone driver: A fine-for-months driver gets replaced by a newer release with a regression, and the creep begins overnight.
- Stuck AI workloads keep allocating: When the NPU appears stuck at zero percent usage while a workload hangs, the driver may keep retrying and allocating indefinitely.
How to fix it
Work through these in order — diagnose first, then act.
Fix 1: Prove the leak with Task Manager and Performance Monitor
Never chase an unmeasured leak: an hour of observation separates a genuine driver leak from ordinary high memory use.
- Open Task Manager’s Details tab, right-click the columns, and add Handles, Memory (active private working set), and Commit size.
- Identify the NPU driver’s host process (often a system service host or the vendor’s NPU service) and note its starting values.
- Use AI features normally for an hour — video calls with effects, Recall browsing — then compare. Steady growth that never retreats after closing everything is a leak.
- For hard evidence, run Performance Monitor, add that process’s Process counters, and log for a few hours: a staircase graph is the textbook leak signature.
Fix 2: Identify which workload triggers the worst leaking
Not all AI features leak equally. Knowing the trigger lets you work around it while you fix the driver, and tells you exactly what to test after each fix.
- Reboot for a clean baseline and record the driver’s memory figure.
- Run one AI feature at a time for 20 minutes each — Studio Effects, live captions, Recall — noting memory after each.
- Note which feature grows the number fastest — Recall with a full snapshot store is a frequent offender thanks to its long indexing sessions.
- Avoid or limit that feature for now; re-test it after each driver fix below.
Fix 3: Update to the latest OEM NPU driver
Most leaks are fixed in later driver releases — vendors find them through the telemetry you just gathered. The OEM package for your exact model is the safest upgrade path.
- Find your exact model in Settings > System > About.
- On Dell , HP , or Lenovo support, download the newest NPU/chipset driver.
- Install it over the current driver (leaks rarely need a purge first), restart, and repeat the one-hour measurement from Fix 1.
- If the growth is gone or dramatically reduced, you’re done — keep this driver and guard it from Windows Update.
Fix 4: Block Windows Update from reinstalling the leaky driver
The cruelest version: you install the fixed driver, Windows Update replaces it with the leaky one, and the creep returns. If your leak started after a Patch Tuesday, this is likely what happened.
- In Device Manager, open the NPU’s Properties > Driver and check the driver date against the version you installed.
- If Windows replaced your driver, Roll Back Driver, then reinstall your known-good OEM package.
- Prevent recurrence with the device-ID installation restriction in gpedit.msc (Computer Configuration > Administrative Templates > System > Device Installation > Device Installation Restrictions), adding the NPU’s hardware ID from Details > Hardware Ids.
- After the next Windows Update cycle, verify your driver version is still in place.
Fix 5: Reduce the leak’s fuel — trim background AI workloads
The driver is the root cause, but reducing AI workload volume slows the leak dramatically and can keep a machine usable until a fixed driver ships.
- Open Settings > Privacy & security > Recall & snapshots: lower snapshot frequency or pause Recall during long sessions.
- Turn off unneeded Studio Effects — each active effect is another allocation cycle.
- In Task Manager’s Startup apps, disable unused vendor AI utilities; some leak slowly even at idle.
- As a stopgap, schedule a nightly 4 AM restart (Task Scheduler,
shutdown /r /t 0) so leaked memory never accumulates past one day.
Fix 6: Do a clean driver reinstall when updates don’t stop the leak
If the newest OEM driver still leaks, the installed copy may be corrupt — a half-written install can leak where a clean one doesn’t. A full purge plus fresh install is the last software step before escalation.
- In Device Manager, uninstall the NPU, checking “Attempt to remove the driver for this device”, then restart.
- Install the latest OEM driver package as administrator and restart.
- Re-run the Fix 1 measurement protocol for at least two hours of real AI use.
- If the leak survives a clean install of the newest driver, report it to your laptop maker’s support with your measurements — driver version, triggering workload, and growth rate per hour. That report is what gets leaks fixed in the next release.
Closing
An NPU memory leak announces itself as creeping slowness a reboot temporarily cures. Measure first, find the triggering workload, update to the latest OEM driver, and protect it from Windows Update. The leak is a software defect, not a reason to buy more RAM — the right driver flattens your memory graph.
