Worklets Legacy Eval Mode memory consumption regression in production

[SirJalias]

Description

On React Native 0.85.3 (New Architecture, Hermes V1), simply importing react-native-reanimated — without rendering a single animated component or calling any of its APIs — increases native RAM usage by ~25–30% compared to RN 0.83.x. Removing the import alone restores the baseline memory profile.

This was originally filed as a React Native memory regression and then investigated as a Hermes V1 regression, but the evidence in both threads points back to Reanimated as the trigger. Filing here so the Reanimated team has visibility.

Cross-references

• React Native: [Android] - Higher memory usage on 0.85.x vs 0.83.x with the new Hermes (~25–30% more RAM, same app) react/react-native#57059 — [Android] - Higher memory usage on 0.85.x vs 0.83.x with the new Hermes (~25–30% more RAM, same app)
• Hermes: Severe memory regression in V1 facebook/hermes#2048 — Severe memory regression in V1
  • In particular this comment: Severe memory regression in V1 facebook/hermes#2048 (comment)

if you simply remove reanimated import you will see normal baseline for rn app
however just import of reanimated without any usage of it cause this problem
I know we can't go without reanimated but from my point of view this is not hermes v1 problem

Environment

react-native	0.85.3
react-native-reanimated	4.4.0
react-native-worklets	0.9.1
react	19.2.3
Architecture	New Architecture (Fabric + TurboModules)
JS engine	Hermes V1 (bytecode compiled)
Platform	Android
Device	Pixel 9 Pro, Android 16
Build	Release / production build

Previously-working baseline: react-native 0.83.9 (Hermes < V1).

Observed memory (production build, measured with Flashlight)

Phase	RN 0.83.9	RN 0.85.3	Delta
Startup RAM	~370 MB	~480 MB	+110 MB
End-state RAM	~560 MB	~720 MB	+160 MB

Heap inspection on Hermes V1 shows most of the extra allocation coming from repeated 256 KB blocks originating in Hermes — which only appear once Reanimated is in the bundle.

Why we believe this is on the Reanimated side

• The RN-only and Hermes-only investigations both reproduce the regression, but the variable that flips it on/off is the Reanimated import, not app code or animation usage.
• The Hermes maintainers' thread (Cant find variable _setGlobalConsole_f with expo #2048) concludes Hermes V1 is expected to use less RAM, and the reporter narrowed the trigger to the Reanimated import.
• Removing the import — with everything else (RN 0.85.3, Hermes V1, New Arch) unchanged — restores the baseline.

Steps to reproduce

1. App on RN 0.83.9 — build a release APK, run a standardized UI flow (FlashList scrolling, screen navigation, repeated interactions) via Maestro, profile RAM with Flashlight. Record baseline.
2. Upgrade to RN 0.85.3 (Hermes V1, New Architecture) with no application code changes. Repeat the same flow → ~25–30% higher RAM.
3. Isolation step: in the 0.85.3 build, remove the import 'react-native-reanimated' (and any reanimated imports) without otherwise changing the app. → Memory returns to the 0.83.x baseline.

The key signal is step 3: the regression is present from the mere presence of the import, before any animation is mounted or any worklet runs.

Snack or a link to a repository

https://github.com/SirJalias/AwesomeProject

Reanimated version

4.4.0

Worklets version

0.9.1

React Native version

0.85.3

Platforms

Android

JavaScript runtime

Hermes

Workflow

React Native CLI

Architecture

New Architecture (Fabric renderer)

Reanimated feature flags

None

React Native release level

Stable

Build type

Debug app & production bundle

Device

Real device

Host machine

macOS

Device model

Pixel 9 Pro, Android 16

Acknowledgements

Yes

[mvrty]

This happens on iOS as well.

[rodperottoni]

I'll cross-post my comment from one of the linked issues:

On further inspection, a production iOS build seems to consume 4x the amount of RAM in Hermes V1. In the images below, RAM usage for our production app went from ~60mb to ~240mb during startup.

Hermes < v1	Hermes v1

[tjzel]

We are currently investigating this, thanks for the issue, reports and reproductions, they are really helpful.

@rodperottoni Just to make sure we're on the same page, the measurements for memory usage were made on the same RN version, just with Hermes V1 enabled and disabled?

[rodperottoni]

We are currently investigating this, thanks for the issue, reports and reproductions, they are really helpful.

@rodperottoni Just to make sure we're on the same page, the measurements for memory usage were made on the same RN version, just with Hermes V1 enabled and disabled?

Hey, the results were from the same app but running Expo 55 vs 56. No code changes besides the package upgrades.

[SirJalias]

@tjzel on my side I just switched the React native version, that is the same behavior on Expo, as v55 used 0.83 and the v56 switched to v.85

[unendingblue]

@tjzel would you be able to provide us with a progress update on this issue? disabling HermesV1 on Expo seems to be quite flaky in our tests so we're kinda stuck. do you reckon this will be resolved soon or is it a matter of weeks?

[zzz08900]

@tjzel would you be able to provide us with a progress update on this issue? disabling HermesV1 on Expo seems to be quite flaky in our tests so we're kinda stuck. do you reckon this will be resolved soon or is it a matter of weeks?

AFAIK there are multiple ongoing issues related to Hermes V1 + reanimated/worklets, it might be a good idea to stay with 0.83 or lower until things get neater.

[tjzel]

After double-checking with the Hermes team, I can finally share our conclusions.

What

The issue lies in the new implementation of eval on Hermes V1, which stores some extra debug metadata for eval invocations that return a function — unfortunately, in production builds too. This directly affects the Legacy Eval Mode of react-native-worklets, adding around 512KB of memory for each unique worklet function used at runtime.

Solutions

The Hermes team has assured me they will ship a fix dropping this extraneous metadata to the Hermes version distributed alongside React Native. However, they made no promises about when that would happen. It's a priority for them, but it could still take a couple of weeks.

Workarounds

There are several workarounds available, although none of them simple. I'll rate them from best to worst based on my personal opinion.

Enabling Bundle Mode in your app

Worklets Bundle Mode doesn't depend on eval and is objectively a more performant and safer way of using worklet functions in your application. With today's 0.10.0 release it's considered no longer experimental.

Reanimated versions compatible with Worklets 0.10.0 will be released today too.

Disabling Hermes V1

Hermes V0 doesn't add the aforementioned metadata to eval function results.

Using Hermes compiled from source

Requires cherry-picking the actual fix into the Hermes version distributed with React Native.

Notes

In parallel, I'm experimenting with opt-in behavior for Legacy Eval Mode, that would use Hermes bytecode directly in production builds.
I'll try to ship it as soon as possible, but I can't guarantee it will be applicable to all application setups.