Retaining image uniqueness with a 10,000-image system?

There are actually several questions nested in this thread, so I’ll try to present them in some logical order.

I’ve periodically seen discussions about potential 10,000-image systems, but I’ve always wondered if it would even be possible to preserve image uniqueness with a list so large. I know that there are also semantic elements to memories, but I’ve always considered visual uniqueness to be very important. Having recently finished completing a 2704-image list, the idea of creating a 10,000-image list while retaining uniqueness seems absolutely impossible. In consequence, it seems like the semantic elements would have to play a larger role, which I suspect would result in more time spent per image.

A 10,000-image system would allow for 33% greater efficiency (compared to a 3-digit system) for decimals and a 20% greater efficiency for binary (4 segments, each 0 - 7, 000-000-000-000 - 111-111-111-111), but due to the loss of uniqueness, I expect that the actual benefit would be less. Moreover, for cards–where there is no benefit–I expect that times would actually be a little worse (due to the reasons stated earlier).

I’m curious about what others have to say. Is it actually possible to retain a reasonable level of visual uniqueness with a 10,000-image system? Are the semantic elements strong enough to make this a non-issue? Would you expect there to be a slower rate (images per time)?

Most importantly, is there anybody who actually uses a 10,000-image system? If so, what are his/her times like?

Notes: I think the card problem could be fixed by carefully constructing the card subset with a strict uniqueness condition. There would be no similar-looking images within this subset, so there would be no possible confusion when memorizing cards.

The phonetics underlying a system like this I think would work best with a category-based approach. From experience, though, a category-based approach yields inanimate objects almost exclusively (Or is it just me?), so the list could easily have a disproportionately high number of inanimate objects, making interactions within loci more difficult.


Simon Reinhard uses a 10,000 image system, and he is one of the fastest memorizers in the world. I’m not sure if there are others who use a system like that in practice.

I haven’t made a 10,000 image system, but I speculated about how one might work on these pages:

Here’s how 10,000 images might look using my sound assignments:

We can organize forum posts about these kinds of systems with this tag: #10000-image-systems.


That’s absolutely fascinating! I didn’t know that somebody was actually using a 10,000-image system. I read through all the links you listed. It was very informative, but it sounds like not a lot is actually known about Simon’s system specifically. Has he never shared it? I’m extremely curious. I have so many questions. Do you know how long it took him to learn it? Did he learn the system in segments like you suggested? Do you know if his images are mostly people, objects, or a mix? I’d really like to contact him if possible. I’m astonished that he actually got a 10,000-image system to work, and very well apparently. That’s incredible.

I don’t think he has ever shared the details publicly, but you could search through his posts here.

Well if Simon’s system is posted anywhere, it’s extremely well hidden. I spent a lot of time searching, but I couldn’t find anything. I also saw several posts from you, Josh (from 2010/2011). In one you actually asked Simon about the details of his system, but he only responded with a few words, basically stating that the 4-digit system exists, but that’s it. He seems reluctant to share any more.

I also came across a post (also from you), where it was mentioned that Ben had started making a 4-digit system, too. Ben’s approach was to form two-part names, each consisting of a vowel and a consonant. This is especially interesting for me, because it may mean that there’s a way to adapt the Ben system to a 10,000 image system later on, making use of the images that I already know. Of course, some of the phonemes wouldn’t carry over completely, but these could just be remapped according to some sort of rule, therefore keeping all 2704 images.

Do you know anything about Ben’s progress by chance? All these posts are quite old.

1 Like

I’m not sure, but he might have written about it on his blog.

1 Like

Yeah, I’m going to have to go through that. I could probably do it most efficiently by searching for key words. Hopefully it yields something. Regardless, even if there were a 10,000 image system I could refer to, I think continuing with my current system would be prudent for now. It’d be nice, though, to have an option of expansion down the road. We’ll see. Maybe a well-trained 3-digit system can be just as effective. Just look at some of the times with 52 images.

1 Like

I was thinking more today about a potential 10,000-image system, and I think I know how I’d do it. Here are my thoughts:

It would be vowel-consonant-vowel-consonant and category based. The first vowel-consonant segment would correspond to a category, and the second segment would correspond to the element in that specific category. There would then be 100 categories, each with 100 elements. This sort of separation would allow for easily associating images with phonemes. Here’s an example:

5963 would be lO (category) + gi (element) → lO is Lord of the Rings, and gi within this subset is Gimli, so the image corresponding to 5963 would be Gimli. This sort of approach directs your brain to the necessary subset, so you’re just left with selecting an element from a list of 100.

If you wanted to start with a 0 - 999 system, you could just start by filling out all elements for the categories that have leading zeros, which would be s x (10 vowels). So you’d be starting by filling out 10 100-element categories.

edit This idea is extremely intriguing. You could start with a 3-digit decimal system (512 binary) and then gradually make additions of 1000. I would begin by filling out all categories, then I’d do 0 - 999, test the waters, and then make a decision about whether to advance or not.

Here’s a quick screenshot of what the first step might look like.

1 Like

The 10 x 10 x 10 x 10 idea works beautifully for decimal and binary, but it poses problems for cards. Even though this system would contain over 3 times the required 2704, the lack of columns with 13 values (for indices) creates a roadblock.

Any ideas on how to sidestep this issue without compromising the decimals or binary? Is it even possible?

I can think of one solution, but it would affect speed. From left to right the columns would be suit-index-suit-index. Each suit would be mapped to 2 numbers depending on the card index. For indices 1 (ace) through 10, you would use one number of the possible two for the suit, and for the remaining indices you would use the other. This seems very slow, though. Is there another way?

edit Actually I think index-suit-index-suit would be better. The index value would be fixed and then it would determine which number the suit would be assigned to.

edit again I guess it wouldn’t be too hard. Whenever you see a face card you simply execute the alternate “block” for suits. I believe it’s this sort of decision process that Alex uses in his block system, so there is a precedent for it not affecting speed.

1 Like

Wow! Mind boggling.
James P.

This is an interesting discussion, particularly with regards to trying to understand the actual payoff of learning a 10,000+ image system.

You can also take a look at the phonetic system I developed earlier this year for all 1 through 4-digit numbers. It’s also based on previous work of other members on this forum. Ben seemed to like the result as well.

Unfortunately the very next day after completing this system I was sidetracked with the need to learn Mandarin as quickly as possible. However I still hope to fill out this system at a more convenient time, at which point I’ll at least be able to draw on a large set of Mandarin words for possible number-image candidates.

I’ll be interested to hear more of how your progress goes!


Slate, I checked out your system and I have to say that I’m incredibly impressed. You did an outstanding job. Your system feels very similar to the Ben system, and all monosyllabic too. Also, something I’m particularly excited about is how easily many of my current image-phoneme pairs could transfer over. If I wanted to, within a week I could have a working 3-digit system, along with a 512 binary system. No cards, though. For that I would have to use a 52-image system.

This has given me a lot to think about.


I was thinking more about a 4-digit system yesterday and I came up with a way to adapt the Ben system to a four-digit system. It’s based on the category approach I described earlier.

All numbers are directly incorporated by being preceded by a zero. For instance, 104 becomes 0104. So whenever a four-digit segment begins with a zero, the following three digits are just pulled from the Ben system. In addition to starting the fill-out process at 10%, you get to keep a monosyllabic system for all numbers 0 - 999.

But there’s a way to adapt the Ben system further by using some of the other monosyllabic phonemes not used in the numbers. The initial and final consonants (or combinations) can be used to group numbers 10 - 15 and 10 - 12 respectively. No number could contain two of these consonants as that would produce a 5-digit number.

There are six of these leading consonants and three final consonants:
initial: d, p, h, sk, st, sh
final: d, p, j

Images added from the six leading consonants: 6 x 100 = 600
Images added from the three final consonants (only when leading number is 2 - 9): 8 x 10 x 3 = 240
Total images transferred = 1000 + 600 + 240 = 1840, or 68%

And this would obviously start you at 18.4% completeness, which is pretty good in my opinion. The only concern that I have though is binary. The spread is so great that these sorts of alternate rules could get very tricky. Maybe I should just keep the leading consonants.

edit I should add that the consonants common to both groups (initial and final) would have the same values. For instance, you could do something like this:
p = 10
d = 11
And then maybe j would be paired with h and both would = 12.

I thought about this some more, and here’s what I’ve come up with: While both contingencies could easily be applied to decimals, the contingency for the final consonants would absolutely cause problems. The conditions for the leading consonants would be fine though. In fact, it works beautifully. Because all leading consonants in the Ben system are mapped to a binary value, the highest possible value, 15, corresponds to 1111, so all you need to do is look for any scenario where you have two leadings 0’s in any 12-digit binary segment. This even allows you to retain the same binary rule set, so if you’re been using the Ben system, the transition would be very easy.

In light of this, you would only be able to retain 1600 images, which is 59%, but the transition would be seamless.

I’ll likely start to chip away at this project after I graduate, and in the meantime I might set up a thread so I can get help brainstorming potential categories.

1 Like