AngularJS remembers the value and compares it to a previous value. This is basic dirty-checking. If there is a change in value, then it fires the change event.

The $apply() method, which is what you call when you are transitioning from a non-AngularJS world into an AngularJS world, calls $digest(). A digest is just plain old dirty-checking. It works on all browsers and is totally predictable.

To contrast dirty-checking (AngularJS) vs change listeners (KnockoutJS and Backbone.js): While dirty-checking may seem simple, and even inefficient (I will address that later), it turns out that it is semantically correct all the time, while change listeners have lots of weird corner cases and need things like dependency tracking to make it more semantically correct. KnockoutJS dependency tracking is a clever feature for a problem which AngularJS does not have.

Issues with change listeners:

• The syntax is atrocious, since browsers do not support it natively. Yes, there are proxies, but they are not semantically correct in all cases, and of course there are no proxies on old browsers. The bottom line is that dirty-checking allows you to do POJO, whereas KnockoutJS and Backbone.js force you to inherit from their classes, and access your data through accessors.
• Change coalescence. Suppose you have an array of items. Say you want to add items into an array, as you are looping to add, each time you add you are firing events on change, which is rendering the UI. This is very bad for performance. What you want is to update the UI only once, at the end. The change events are too fine-grained.
• Change listeners fire immediately on a setter, which is a problem, since the change listener can further change data, which fires more change events. This is bad since on your stack you may have several change events happening at once. Suppose you have two arrays which need to be kept in sync for whatever reason. You can only add to one or the other, but each time you add you fire a change event, which now has an inconsistent view of the world. This is a very similar problem to thread locking, which JavaScript avoids since each callback executes exclusively and to completion. Change events break this since setters can have far-reaching consequences which are not intended and non obvious, which creates the thread problem all over again. It turns out that what you want to do is to delay the listener execution, and guarantee, that only one listener runs at a time, hence any code is free to change data, and it knows that no other code runs while it is doing so.

What about performance?

So it may seem that we are slow, since dirty-checking is inefficient. This is where we need to look at real numbers rather than just have theoretical arguments, but first let's define some constraints.

Humans are:

• Slow — Anything faster than 50 ms is imperceptible to humans and thus can be considered as "instant".

• Limited — You can't really show more than about 2000 pieces of information to a human on a single page. Anything more than that is really bad UI, and humans can't process this anyway.

So the real question is this: How many comparisons can you do on a browser in 50 ms? This is a hard question to answer as many factors come into play, but here is a test case: http://jsperf.com/angularjs-digest/6 which creates 10,000 watchers. On a modern browser this takes just under 6 ms. On Internet Explorer 8 it takes about 40 ms. As you can see, this is not an issue even on slow browsers these days. There is a caveat: The comparisons need to be simple to fit into the time limit... Unfortunately it is way too easy to add a slow comparison into AngularJS, so it is easy to build slow applications when you don't know what you are doing. But we hope to have an answer by providing an instrumentation module, which would show you which are the slow comparisons.

It turns out that video games and GPUs use the dirty-checking approach, specifically because it is consistent. As long as they get over the monitor refresh rate (typically 50-60 Hz, or every 16.6-20 ms), any performance over that is a waste, so you're better off drawing more stuff, than getting FPS higher.

Notice: I do update this answer as I find better solutions. I also keep the old answers for future reference as long as they remain related. Latest and best answer comes first.

Better answer:

Directives in angularjs are very powerful, but it takes time to comprehend which processes lie behind them.

While creating directives, angularjs allows you to create an isolated scope with some bindings to the parent scope. These bindings are specified by the attribute you attach the element in DOM and how you define scope property in the directive definition object.

There are 3 types of binding options which you can define in scope and you write those as prefixes related attribute.

angular.module("myApp", []).directive("myDirective", function () {
    return {
        restrict: "A",
        scope: {
            text: "@myText",
            twoWayBind: "=myTwoWayBind",
            oneWayBind: "&myOneWayBind"
        }
    };
}).controller("myController", function ($scope) {
    $scope.foo = {name: "Umur"};
    $scope.bar = "qwe";
});

HTML

<div ng-controller="myController">
    <div my-directive my-text="hello {{ bar }}" my-two-way-bind="foo" my-one-way-bind="bar">
    </div>
</div>

In that case, in the scope of directive (whether it's in linking function or controller), we can access these properties like this:

/* Directive scope */

in: $scope.text
out: "hello qwe"
// this would automatically update the changes of value in digest
// this is always string as dom attributes values are always strings

in: $scope.twoWayBind
out: {name:"Umur"}
// this would automatically update the changes of value in digest
// changes in this will be reflected in parent scope

// in directive's scope
in: $scope.twoWayBind.name = "John"

//in parent scope
in: $scope.foo.name
out: "John"


in: $scope.oneWayBind() // notice the function call, this binding is read only
out: "qwe"
// any changes here will not reflect in parent, as this only a getter .

"Still OK" Answer:

Since this answer got accepted, but has some issues, I'm going to update it to a better one. Apparently, $parse is a service which does not lie in properties of the current scope, which means it only takes angular expressions and cannot reach scope. {{,}} expressions are compiled while angularjs initiating which means when we try to access them in our directives postlink method, they are already compiled. ({{1+1}} is 2 in directive already).

This is how you would want to use:

var myApp = angular.module('myApp',[]);

myApp.directive('myDirective', function ($parse) {
    return function (scope, element, attr) {
        element.val("value=" + $parse(attr.myDirective)(scope));
    };
});

function MyCtrl($scope) {
    $scope.aaa = 3432;
}​

.

<div ng-controller="MyCtrl">
    <input my-directive="123">
    <input my-directive="1+1">
    <input my-directive="'1+1'">
    <input my-directive="aaa">
</div>​​​​​​​​

One thing you should notice here is that, if you want set the value string, you should wrap it in quotes. (See 3rd input)

Here is the fiddle to play with: http://jsfiddle.net/neuTA/6/

Old Answer:

I'm not removing this for folks who can be misled like me, note that using $eval is perfectly fine the correct way to do it, but $parse has a different behavior, you probably won't need this to use in most of the cases.

The way to do it is, once again, using scope.$eval. Not only it compiles the angular expression, it has also access to the current scope's properties.

var myApp = angular.module('myApp',[]);

myApp.directive('myDirective', function () {
    return function (scope, element, attr) {
        element.val("value = "+ scope.$eval(attr.value));
    }
});

function MyCtrl($scope) {
   
}​

What you are missing was $eval.

http://docs.angularjs.org/api/ng.$rootScope.Scope#$eval

Executes the expression on the current scope returning the result. Any exceptions in the expression are propagated (uncaught). This is useful when evaluating angular expressions.