Skip to main content

Memoizing metaclass. Drop-dead simple way to create cached objects

Project description

A quick way to make Python classes automatically memoize (a.k.a. cache) their instances based on the arguments they are instantiated with (i.e. args to __init__). It’s a simple way to avoid repetitively creating expensive-to-create objects, and to make sure objects that have a natural ‘identity’ are created only once.


Say you have a class Thing that requires expensive computation to create, or that should be created only once. In Python 2.x you can make that happen by adding one line to its definition:

from mementos import MementoMetaclass

class Thing(object):

    __metaclass__ = MementoMetaclass  # now I'm memoized!

    def __init__(self, name): = name


Then Thing objects will be memoized:

t1 = Thing("one")
t2 = Thing("one")
assert t1 is t2    # same instantiation args => same object

Python 3

mementos works in Python 3 just as in Python 2, but with different syntax. Instead of the double-underscore class attribute assignment, Python 3 uses a keyword argument at class creation:

class Thing3(object, metaclass=MementoMetaclass):

Unfortunately, Python 2 and Python 3 don’t recognize each other’s syntax for metaclass specification, so straightforward code for one won’t even compile for the other. You can get around this by using the with_metaclass() function, which is similar to that found in the six cross-version compatibility module.:

from mementos import MementoMetaclass, with_metaclass

class Thing23(with_metaclass(MementoMetaclass, object)):

Careful with Call Signatures

MementoMetaclass caches on call signature, which can vary if keyword args are used. E.g. def func(a, b=2) could be called func(1), func(1,2), func(a=1), func(1, b=2), or func(a=2, b=2)–and all resolve to the same logical call. And this is just for two parameters! If there are more than one kwarg, they can be arbitrarily ordered, creating many logically identical permuations. Thank Goodness Python doesn’t allow kwargs to come before positional args, else there’d be even more ways to make the same call.

So if you instantiate an object once, then again with a logically identical call but using a different calling structure/signature, the object won’t be created and cached just once–it will be created and cached multiple times.:

o1 = Thing("lovely")
o2 = Thing(name="lovely")
assert o1 is not o2     # because the call signature is different

This may degrade performance, and can also create errors, if you’re counting on mementos to create just one object. So don’t do that. Use a consistent calling style, and it won’t be a problem.

In most cases, this isn’t an issue, because objects tend to be instanitated with a limited number of parameters, and you can take care that you instantiate them with parallel call signatures. Since this works 99% of the time and has a simple implementation, it’s worth the price of this inelegance.

If you want only part of the initialization-time call signature (i.e. arguments to __init__) to define an object’s identity/cache key, there are two approaches. One is to use MementoMetaclass and design __init__ without superflous attributes, then create one or more secondary methods to add/set useful-but-not-essential data. E.g.:

class OtherThing(with_metaclass(MementoMetaclass, object)):

    def __init__(self, name): = name
        self.color = None   # unset for now
        self.weight = None

    def set(self, color=None, weight=None):
        self.color = color or self.color
        self.weight = weight or self.weight
        return self

ot1 = OtherThing("one").set(color='blue')
ot2 = OtherThing("one").set(weight='light')
assert ot1 is ot2
assert ot1.color == ot2.color == 'blue'
assert ot1.weight == ot2.weight == 'light'

Or you can just define your own memoizing metaclass, using the factory function described below.

Visiting the Factory

The first iteration of mementos defined a single metaclass. It’s since been reimplemented as a parameterized meta-metaclass. Cool, huh? That basically means that it defines a function, memento_factory() that, given a metaclass name and a function defining how cache keys are constructed, returns a corresponding metaclass. MementoMetaclass is the only metaclass that the module pre-defines, but it’s easy to define your own memoizing metaclass.:

from mementos import memento_factory, with_metaclass

IdTracker = memento_factory('IdTracker',
                            lambda cls, args, kwargs: (cls, id(args[0])) )

class MyTracker(with_metaclass(IdTracker, object)):

    # object idenity is the object id of first argument to __init__
    # (and there must be one, else the args[0] reference => IndexError)

The first argument to memento_factory() is the name of the metaclass being defined. The second is a callable (e.g. lambda expression or function object) that takes three arguments: a class object, an argument list, and a keyword arg dict. Note that there is no * or ** magic–args passed to the key function have already been resolved into basic data structures.

The callable must return a globally-unique, hashable key for an object. This key will be stored in the _memento_cache, which is a simple dict.

When various arguments are used as the cache key/object identity, you may use a tuple that includes the class and arguments you want to key off of. This can also help debugging, should you need to examine the _memento_cache cache directly. But in cases like the IdTracker above, it’s not mandatory that you keep extra information around. The raw id(args[0]) integer value would suffice, as would a constructed string or other immutable, hashable value.

In cases where arguments are very flexible, or involve flexible data types, a high powered hashing function such as that provided by SuperHash comes in extremely handy. E.g.:

from superhash import superhash

SuperHashMeta = memento_factory('SuperHashMeta',
                            lambda cls, args, kwargs: (cls, superhash(args)) )

For the 1% edge-cases where multiple call variations must be conclusively resolved to a unique canonical signature, that can be done on a custom basis (based on the specific args). Or in Python 2.7 and 3.x, the inspect module’s getcallargs() function can be used to create a generic “call fingerprint” that can be used as a key. (See the tests for example code.)


  • mementos is not to be confused with memento, which does something completely different.

  • Mementos was originally derived from an ActiveState recipe by Valentino Volonghi. While the current implementation quite different and the scope much broader, the availability of that recipe was what enabled this module and the growing list of modules that depend on mementos. This is what open source evolution is all about. Thank you, Valentino!

  • It is safe to memoize multiple classes at the same time. They will all be stored in the same cache, but their class is a part of the cache key, so the values are distinct.

  • This implementation is not thread-safe, in and of itself. If you’re in a multi-threaded environment, consider wrapping object instantiation in a lock.

  • Automated multi-version testing managed with pytest and tox. Successfully packaged for, and tested against, all late-model verions of Python (2.6, 2.7, 3.2, and 3.3), plus one (2.5) that isn’t so very recent, and one (PyPy 1.9, based on Python 2.7.2) that is differently implemented.


pip install mementos

To easy_install under a specific Python version (3.3 in this example):

python3.3 -m easy_install mementos

(You may need to prefix these with “sudo ” to authorize installation. If they’re already installed, the --upgrade flag will be helpful; add it right before the package name.)

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions (17.9 kB view hashes)

Uploaded source

mementos-1.0.tar.gz (8.6 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page