Well hello there!

Are you tired of dealing with JLLM's antics and shenanigans?, are you tired of ruining/getting ruined for everyone else?, are you tired of getting your body and soul claimed?, tired of getting your earlobe nipped at?, and tired of characters forgetting who you are, what you did and what clothes you have on every 5 messages?.

And on the other hand, are you tired of running out of proxy requests?, do you refuse to pay to get a decent ammount of openrouter messages and you have to scavenge for working models that arent rate-limited?.

Wouldnt you like, to be able to just get the best experience from both worlds?, be able to extend your free proxy requests while also having long, decent roleplays that get derailed and end in the total death of your character's identity and personality?. just becuase you cant talk for more than 50 messages without openrouter sending you to hit hay?.

Well, look no further!, for i have deviced a system that if followed to a T, will guarantee (or at least, do something), to improve your roleplaying experience greatly, and all you need is access to a free proxy, and the enough working neurons to keep count of your messages.

This, is the style anchor and memory loop system.

DISCOVERY

The discovery of this system ocurred one day where i was playing around with a proxy, chatting around, seeing the differences between JLLM responses and The proxy's and marveling at how livid the proxy was and how...serviceable but definitely lacking JLLM's responses were (Just in case anyone is thinking about it, this post is not an insult or complaint about the quality of JLLM, its an amazing model for being 100% free and unlimited and i really appreciate the janitor dev team for making it possible to work for everyone).

However as i was experimenting i discovered something, any JLLM message sent right after a proxy message seemed to get, enhanced, a lot. JLLM (like many other smaller LLMs) seems to have a very good capacity for mimicry, if it gets fed a well written message, it will do its best to mimic its style until it gets deleted from its memory and it goes back to its default state, this applies both to proxy and hand-written messages.

So right there and then i realized that, this could be exploited, i could technically stretch the quality of proxy messsages by having JLLM mimic them, it wouldnt be as good, and would only last until JLLM's memory has to start deleting stuff to keep up, but it would save a ton on proxy requests. So i got hands on and came up with a system that works surprisingly well

WHAT YOU WILL NEED

• JLLM (Everyone has JLLM, it came free with your fucking JAI account)

• Access to a good proxy (prefferably deepseek or anything with more memory than JLLM)

• A way to count your messages (If you cant do this yourself, god bless you)

• Dedication to the craft (this isnt for lazy users)

PRINCIPLE

The main idea of this system is to use a bigger model (like deepseek), to "power-up" JLLM by injecting its message style into the chat's memory, as well as the use of proxy made summaries to create a pseudo-expanded memory that will make JLLM more likely to stay consistent with the event of your roleplay. (But just to be clear, the main goal of this system is to improve message quality, effects on memory can vary on a chat by chat basis)

HOW TO DO IT

Please, pay attention from here on out, becuase this system has to be executed correctly to ensure it works, any mistake in the process can snowball into a mess that will force you to restart your chat, got it?. So take notes, this is gonna be on the exam.

In order to start with this system, you first need to have a proxy ready, i wont tell you what mine is, becuase i dont want it to get flooded, and i advise you dont go around telling people about YOUR proxy, becuase if you are looking at this, chances are, you are just anothet cheap bastard looking for a way to stretch your 50 daily requests as much as you can.

okay, lets get started, these are the steps to this sytem:

1. Configure your generation settings to have a max of 500 to 800 tokens, 800 is the upper limit becuase its about the max ammount you can go before your messages get too long for JLLM to handle properly. so keep them between those two values. I dont care about temperature or whatever. Any time you switch from JLLM to proxy and back, make sure to NOT reset your config back to default.

2. Start your chat with a proxy generated messsage, NOT jllm. This will be your first "anchor message", it will be the style guide JLLM is meant to follow, vocabulary may vary between the proxy and JLLM, but the message format will in most cases, stick quite nicely. (Though vocabulary also changes as more and more anchor messages get introduced)

3. Switch back to JLLM, and keep roleplaying as normal for about 20 messages (10 of yours + 10 bot responses), at this point you may start to notice how JLLM is starting to loose steam, you may see some repetition, messags lengthening, usual JLLM shenanigans,etc

4. Before you sent your 21st message, switch back to the proxy and generate another response, this will create another anchor message, it will reinforce the format, style, vocabulary and personality of the character, JLLM will improve performance again.

5. Switch back to JLLM, and keep roleplaying as normal untl you reach the 40th message, at this point JLLM's memory may start to get very full and forgetful, its time for a clean slate.

6. Switch to the proxy, and ask it to generate a detailed summary of the last 40 messages of the chat (or as far back as it can go), copy and paste the output into the chat memory feature, this will purge JLLMs memory and replace it with the recent events on the roleplay, then once the summary is placed in the chaat memory, delete YOUR message asking the proxy for it. The reason for that is that JLLM usually takes cue of the prior message as a basis for the next one, so if you leave the summary in the chat, JLLM will start writing in the format of the resume, not the rest of the messages, this will also affect the proxy in future cycles so just get rid of it once its pasted on the. Your results may vary depending on the depth and complexity of the roleplay and quality of the summary. Despite saying that you should do this in your 40th messages, more complex and heavy roleplays may benefit from summaries every 20 or even 10 messages. So if the summaries dont work for you, try making them more often, or just dont make them at all. Also, dont forget to overwrite old summaries with new ones you make. JLLM can only remember so much, this may make you loose very old things, but who wants to reference what happened 200 messages ago anyways?.

7. Repeat ad infinium, keep counting your messages in "cycles" of 20 messages, every 20 messages send one proxy message, and every 2-3 cycles make a summary, if you fail to do this, the JLLM may introduce character drift and errors that will be very hard to fix, even with a proxy (since it will inherit the mistakes of JLLM).

BONUS (manual mode): Technically, you dont even need to have a proxy for this system to work, as long as you are a good writer, you can just edit the bot's first response into something you like, and as long as you keep introducing your own "anchors" it will mimic and keep up with your style and format. you can also hand-write your summaries if you want.

And thats it!, thats the system, you just need to keep it going while you roleplay and i guarantee, it will make your experience at least a bit better, and it will save on proxy requests.

WHAT THIS SYSTEM DOES AND WHAT IT DOESNT

what it does:

• Improve's JLLM's vocabulary slightly and introduces a strong format style based on the proxy's style

• Slightly improve long term memory in the form of summaries

• Allows for JLLM to work better with token heavy bots without collapsing

• Allows you to reduce your proxy request usage by about 90-95%, turning 50 daily request into aproximately more than 2000 daily messages.

What it DOESNT

• Improve JLLM's vocabulary completely, you will still find some repetition and shenanigans, just less often

• Improve JLLM's performance permanently, you need to keep the anchors and summaries comming for the system to keep working properly

• Improve JLLM's working memory, JLLM will still have a 2K memory, it will still struggle to reference old events during the 20 message cycles, but each anchor and summary may allow you to reference very old things, so if you want a character to mention or talk about something that you know JLLM may not remember, do it in a proxy message or wait till the summary inmortalizes it

• Work 100% of the time, sometimes the proxy may mess up, sometimes JLLM may mess up, maybe the bot is written wierd, maybe you just suck at chatting. lots of variables neither you or i can control can make this system fail. So dont come complaining when it doesnt work for you.

LAST WORDS

"If you do the thing, and you do it right, and you dont it up. It works, it just works"

-jontron 2017

Anyways, im off to play doom.