Introducing the `memoisetools` package

memoisetools package

memoisetools is a collection of additional caches and helper functions to work alongside the memoise package.

This package introduces new caches, new memoise() alternatives and functions for interrogating caches and expiring old objects from a cache.

  • New caches:
    • cache_filesystem2() - with object timestamping, compression of objects by default, and expiration of results not accessed for a certain time.
    • cache_memory2() - with object timestamping, faster xxhash64 used by default, and expiration of results not accessed for a certain time.
  • New memoise::memoise() alternatives
    • memoise_with_result_size_limit() - only store results below a certain size in the cache
    • memoise_with_mixed_backend() - have 2 caches in a memoised function, with small results saved in the first cache, and large objects saved in the second cache.
  • Helper functions
    • get_memoise_info() - to print and return information about the cache(s) of a memoised function e.g. how many objects, total size, etc
    • expire_cache() - If the cache for a memoised functionhas a timestamp, then this function will deleted cached results older than the specified age




get_memoise_info() returns information about the caches used by a memoised function.

  • cache - Cache type. Either ‘memory’, ‘filesystem’ or ‘gcs or aws’
  • Storage location
    • env - For ‘memory’ cache, the R environment in which objects are stored
    • path - For ‘filesystem’ caches, the path to the cache directory
    • cache_name - For ‘gcs or aws’ caches, the name of the cache
  • algo - Hashing algorithm for creating keys
  • bytes - Size (in bytes) of the cache
  • n - Number of objects in the cache
  • has_timestamp - Does the cache include timestamp information?
  • compress - Value for ‘compress’ variable

Note: because memoised functions could have multiple caches (e.g. memoise_with_mixed_backend), this function returns a list of info for each cache.

memoised_rnorm <- memoise::memoise(rnorm)

x <- memoised_rnorm(1000)
y <- memoised_rnorm(12)
z <- memoised_rnorm(1)

Registered S3 method overwritten by 'pryr':
  method      from
  print.bytes Rcpp
cache: memory, env: <environment>, algo: sha512, bytes: 9728, n: 3, has_timestamp: FALSE, compress: FALSE


This is a replacement for memoise::cache_filesystem() with the following changes:

  • use a tempdir() if no path specified
  • Full absolute path to cache is used, even if initialised with a relative path.
    This avoids issues as detailed in this memoise issue on github
  • By default objects saved to filesystem are compressed using gzip compression. (The corresponding memoise PR on github)
  • A separate data structure keeps track of the time of all reads/writes to the cache.
  • The addition of timestamps allows for expiring objects older than a certain age. See the function memoisetools::expire_cache()
  • the cache has_key method now uses the timestamp cache to determine if a given key exists or not. This makes it faster to check if a key exists (as no filesystem access is needed), but will cause an error if they file doesn’t actually exist e.g. if you’ve deleted the file manually.


This is a replacement for memoise::cache_memory() with the following changes:

  • A separate data structure keeps track of the time of all reads/writes to the cache.
  • The addition of timestamps allows for expiring objects older than a certain age. See the function memoisetools::expire_cache()
  • Use the faster hash xxhash64 by default

Expiring objects from the cache

With cache_filesystem2() and cache_memory2(), objects older than a specified age can be retired from the cache. I.e. if they have not been read or written more recently than the specified time, they will be deleted.

memoised_rnorm <- memoise::memoise(rnorm, cache = memoisetools::cache_memory2()) 

memoised_rnorm(1) # stored in cache. 
[1] -1.450964
memoised_rnorm(2) # stored in cache
[1]  0.3509097 -0.1745469
Sys.sleep(1)      # wait a little bit
memoised_rnorm(1) # recent access to this cached data means it won't be expired
[1] -1.450964

# The following expiry will only delete the cached result for `memoised_rnorm(2)`
# as it has not been read/written in over 1 second
memoisetools::expire_cache(memoised_rnorm, age_in_seconds = 1, verbose = TRUE)
cache_memory2: Expired 1 objects

memoised_rnorm(1) # this result is still in the cache
[1] -1.450964
memoised_rnorm(2) # this is a fresh result as the cached version was removed
[1] -0.5914285 -1.3340273


This is a replacement for memoise::memoise() which places a limit on how large an object can be before it is no longer stored in the cache (but simply recalculated each time).

By default, memoise::memoise() will store all results regardless of size. This works for the majority of cases where you have enough memory and results are never too large.

For the problem I was working on, the function produced many small results and a few very very large results (greater than 2GB is size). If all the big results were cached I’d run out of memory!

In the following example, results over 1000 bytes will not be cached.

memoised_rnorm <- memoisetools::memoise_with_result_size_limit(rnorm, result_size_limit = 1000)

memoised_rnorm(1) # small enough to cache
[1] -1.097299
memoised_rnorm(1) # getting cached result
[1] -1.097299

head(memoised_rnorm(1000)) # too big to be cached
[1]  2.0361036 -0.3264896  0.7740052  0.7850064  0.7632461  0.2948088
head(memoised_rnorm(1000)) # so each run produces fresh result
[1]  1.4251896 -1.7347494 -0.8374046 -0.7157526  0.3748210  2.3833280


This is an adjusted version of memoise::memoise() which requires two caches to be set, along with a size limit. Objects smaller than the size limit go to the first cache, and objects larger than the size limit go to the second cache.

This allows you to cache small results in memory, and send large results to the filesystem, s3 or google cloud storage.

memoised_rnorm <- memoisetools::memoise_with_mixed_backend(
  cache1 = memoisetools::cache_memory2(),
  cache2 = memoisetools::cache_filesystem2(),
  result_size_limit = 1000

a <- memoised_rnorm(1) # These 3 results cached to memory
b <- memoised_rnorm(2)
c <- memoised_rnorm(3)

x <- memoised_rnorm(1000)  # These 2 results cached to filesystem
y <- memoised_rnorm(2000)

cache: memory, env: <environment>, algo: xxhash64, bytes: 1648, n: 3, has_timestamp: TRUE, compress: FALSE
cache: filesystem, path: /private/var/folders/5p/78cv9fvn4xn_rbgxpx51q5n80000gn/T/Rtmp4L9t5x, algo: xxhash64, bytes: 23315, n: 2, has_timestamp: TRUE, compress: TRUE


  • Once the archive package matures a little, a more extensive suite of compression options would be great i.e. using lz4 and zstd compression.
  • Proper support for Google Cloud Storage and Amazon S3 backends in get_cache_info()
  • More tools for manipulating caches?
    • Delete all objects over a certain size.
    • Delete oldest objects in order to keep the cache under a certain size.
    • Cache the object size as the results are created and cached? (so we don’t recalculate it each time get_cache_info() is called)
    • Move caches between backends e.g. move a filesystem cache to memory.