Memize is a unabashedly-barebones memoization library with an aim toward speed.

Why use Memize?

• 🚀 It's very fast. Implemented as a least recently used (LRU) cache, it's heavily optimized for scenarios where the function is called repeatedly with the same arguments.
• 🔬 It's tiny. It weighs in at less than 0.3kb minified and gzipped, with no dependencies.
• 🔀 It supports common arguments patterns, including multiple arguments and non-primitive arguments (by reference).

Simply pass your original function as an argument to Memize. The return value is a new, memoized function.

function fibonacci( number ) {
	if ( number < 2 ) {
		return number;
	}

	return fibonacci( number - 1 ) + fibonacci( number - 2 );
}

var memoizedFibonacci = memize( fibonacci );

memoizedFibonnaci( 8 ); // Invoked, cached, and returned
memoizedFibonnaci( 8 ); // Returned from cache
memoizedFibonnaci( 5 ); // Invoked, cached, and returned
memoizedFibonnaci( 8 ); // Returned from cache

Using npm as a package manager:

npm install memize

Memize accepts a function to be memoized, and returns a new memoized function.

memize( fn: Function, options: ?{
	maxSize?: number
} ): Function

Optionally, pass an options object with maxSize defining the maximum size of the cache.

The memoized function exposes a clear function if you need to reset the cache:

memoizedFn.clear();

Implemented as a least recently used (LRU), Memize is heavily optimized for scenarios where the function is called repeatedly with the same arguments. In these scenarios, Memize outperformed most other memoization libraries at the time of initial publication.

To learn more about these benchmarks, important caveats, and how to run them yourself, refer to benchmark/README.md.

If you haven't already, feel free to glance over the source code. The code is heavily commented and should help provide substance to the implementation concepts.

Memize creates a last-in first-out stack implemented as a doubly linked list. It biases recent access favoring real-world scenarios where the function is subsequently invoked multiple times with the same arguments. The choice to implement as a linked list is due to dramatically better performance characteristics compared to Array#unshift for surfacing an entry to the head of the list (jsperf). A downside of linked lists is inability to efficiently access arbitrary indices, but iterating from the beginning of the cache list is optimized by guaranteeing the list is sorted by recent access / insertion.

Each node in the list tracks the original arguments as an array. This acts as a key of sorts, matching arguments of the current invocation by performing a shallow equality comparison on the two arrays. Other memoization implementations often use JSON.stringify to generate a string key for lookup in an object cache, but this benchmarks much slower than a shallow comparison (jsperf).

Finally, special care is made toward treatment of arguments due to engine-specific deoptimizations which can occur in V8 via arguments leaking. Order is important here; we only create a shallow clone when necessary, after the cache has been checked, to avoid creating a clone unnecessarily if a cache entry exists. Looking at the code, you'd not be blamed for thinking that dropping the shallow clone would improve performance, but in fact it would slow execution by approximately 60%. This is due to how the lingering arguments reference would carry over by reference ("leaks") in the node's args property. Update: As of November 2019, engine improvements are such that arguments leaking does not have as dramatic an effect. However, my testing shows that the shallow clone still performs equal or better than referencing arguments directly, and as such the implementation has not been revised in order to achieve optimal performance in the most versions of V8.

Copyright 2018-2020 Andrew Duthie

Released under the MIT License.