Increasing working memory

I thought it would be interesting to consider training working memory as a whole. In general I have thought of some useful approaches that would work if this works at all.

1# of those approaches is less guaranteed to work than the other two but none the less I will state them.

  1. Training your general working memory by giving you a wide variation of ‘things’ you would want to remember, by having them not repeat the goal is to stimulate the underlying working memory aspect without introducing your task specific memory. This requires quite a lot of unique data.

  2. Task specific working memory training, quite simply it’s proven that we can get better at specific tasks with specific things, so this approach is simply to remember the appropriate chunks to increase your working memory at the task. To make it more general being able to visualise more symbols and letter combinations would be useful. You would be able to think about things in symbols which doesn’t have much of a downside if your memory for it is large. The downside is simply the amount of learning it would require, but it would be interesting if you didn’t have to learn chunks by adding one on top of whatever you had and instead could find some in-between generalisation. Nevertheless it’s not too difficult to raise this by quite a bit especially if dedicated, you could think of it as the rote of digit systems.

I have noticed at least that my working memory seems higher when more frequently recognised words appear so this applies much the same to language, though learning language chunks can be nearly endless, perhaps learning a few words that translate to the rest and learning the translations and their chunks may shorten it. Regardless I am pretty convinced that #2 works.

The above 2 approaches take no memory techniques what so ever but are supposed to induce training. The 2nd one is more guaranteed to work but is also lengthy.

  1. My third approach would be using mnemonics, but not in the normal way, in short to quickly store what you want to recall and then recall it without the mnemonics (reinforcing what you are forgetting from the memory palace quickly) then to forget the loci and memory palace and maintain this. There is some more meaning to this, for starters I have realised with the shaper system that when I do see somethings representation inside my memory palace or the other way around the recall to it is nearly instant. Testing this on foreign alphabets has also given me a more structured memory of the letters and symbols that would normally take writing or tracing them with spaced repetition. What I am getting at here is that this can be used to bolster #2 if it doesn’t work on its own. In which case it would require a shaper implementation, while also inducing the forgetting.

I once attempted something similar where I wrote sentences in my memory palace and remembered them then tried to look at all the locations at the same time. Visualising all their meanings in a go was very strenuous but was at higher capacity than you would be able to do normally. I got to around 9 (albeit short) sentences before I was hitting some wall (I combined them one by one into a large paragraph), I’m wondering separately whether increasing the count on this may provide some other general benefits.

I think its possible to add some twists to these to make them more potent e.g ,speed,quantity,efficiency, parallel, but the general idea is there.

What do you guys think, has anyone tried any of these things?

I’m likely to be trying them quite soon, starting with the less time consuming approaches to see if they work as expected.

As for a 'why would you increase your working memory’ : There is a question of whether having more working memory actually has unexpected benefits such as an increasing speed which you can only truly answer if you actually increase your working memory. Then there is the following of arguments(verbal) and the display of information at the same time (visual), which with a memory palace is still sequential and not easily interconnected.

Whether truly significant though easily assumable significant is something you can really find out if you have done it.

Summary of my 3 approaches :
All training based on
general-unique data, task-specific chunks, overloading-difficulty using mnemonics.


After testing a large variety of approaches I have noticed 2 types of success.

  1. By employing as expected a form of chunking and 2 something that fills the hole of the chunking.

It would take a long long page to explain it in detail so I will keep my findings more simple.

First I reasoned out a range of things so I set my goal exclusively to the visual working memory capacity and I reasoned through some obvious points on verbal and visual working memory:

  • Language is extremely large
  • Performance on specialised (little encounter) terms are worse
  • visual working memory while complex is not large per ‘maximum’
  • some memorised statements do not impact your current recall as much as ‘random statements’ proving chunking

So I set out to test specific things in visual working memory to enhance the capacity at a very small step. The idea was that if I can far exceed my visual working memory by having ‘a chunk’ present as part of the memory. Then I have in theory proven that this approach functions for working memory and I could test any dependencies after that.

I ran the tests by visualising a 3x3 matrix successfully and then visualising a 2x2 matrix on the right or left of that in my head without any ‘techniques’. Then by considering possibilities I tried this with

  • the method of loci
  • objects
  • shaper locations
  • shaper objects
  • high speed
  • drawing in the digits
  • visualised spaced repetition
  • chunking variations

While I have tried a lot of these variations only truly 2 general things worked. 1. ‘getting the image of all numbers to stay in my head at a time’. 2. drawing the images of the numbers in.

The important discovery here is that essentially the shaper system by placing the objects close together rather than spread out (higher speed also makes this easier) you can effectively get the images of the digits to stay represented in your head. This said there was a significant issue I noticed with this. First it took some getting used to , in order to represent the total digits. Secondly I noticed that it felt very off, it felt like when I am looking at a drawing. Where I can’t really recall the exact precise detail of the drawing but I can clearly remember it. In this case it was sufficient for recall but I had a flash of realisation at this point.

Generally when I look at a page of text and try to recall it, its easy to recall that there was a page of text just as it is, but I then can’t read this page of text. I can see the shapes the text is forming and know exactly that there is text there, but can’t read it. I remembered in some of my language learning times this similar thing was also there with foreign language characters and it went away once I drew the language characters rather than simply learned to recall them. With this in mind I tried drawing in the digits (in my head) and noticed that this odd feeling started vanishing, I had a kind of grasp of the digits after doing this. This gave me a chain of light-bulbs such as why visualised abacus even work.

Essentially however this ‘drawing’ in is essential to capture the detail of the representation, and the final representation then contains the link to the detail (if you have it) and simply a picture of the representation if you do not. This kind of odd intuitive creation of detail made a lot of sense at this point. For example after many years of using the abacus some people achieve very fast speeds, I had wondered what exactly the difference was between a visualised abacus and simply visualising the digits as they are. It was obvious to me that at high speeds people were no longer actually moving the beads, it seemed illogical, they would by then memorise the placements and automatically get to them based on the sum. Up until now this didn’t quite make a lot of logical sense ‘why can’t you simply do this with visualised numbers?’. However when I consider this its true that the detailed operation can simply retain itself in the final representation provided it is a transformation of what is already there. It’s significantly easier with soroban beads.

This realisation has changed a lot of things for me personally, I reasoned out if this was also the case with speech but that aside, it has lead me to a more general approach of enhancing my visual working memory, which is to actively draw simultaneous parts of whatever I am trying to recall at the same time (of course at high speed). Doing this achieves me a decent digit span without any subvocalization and certainly faster than subvocalization it also transfers to other things , though given the amount of letters in language it needs some refining to be used for language. It can easily be used by generating letters one is already used to or combinations of which I have some practice with rather than the individual drawing of them. That said, it is working memory not long term memory though it does transfer easier to long term memory in my opinion. It also simplifies working memory and implies that chunking is effectively the use of short term memory with working memory.

These however are just flashes of experimental reasoning. They kind of change or add on to a lot of the science on this so it can be seen as controversial. My own results however are positive of this.

Either way I have found an interesting new way to improve my visual working memory, and I have gotten a better idea of how to make chunking effectively work as well as use some new interesting ability. What I haven’t yet tested is the intensity approach(1) as that is quite time consuming and tiring.

the drawing can be done with shaper objects too by for example constructing them, this tends to be easier to do but there is no issue with or without mnemonics.

If this property of visual working memory is really valid and I haven’t just made an error somewhere perhaps in the reasoning of theory or practical testing then it’s almost certainly going to be valid for things other than visual working memory.

I will post an update on this if there are any revisions since it’s been around 3 days since I have had success with this or if I run into issues in increasing working memory or otherwise.