The last update was 19 days ago. Don't get me wrong, I just want to know if there will be another update in the near future and if entities will be fixed. Sorry if it sounds rude, English is not my language.
Every choice transforms your body and shapes your adventure · By
While I would also like an update with hopefully some added functionality, the entities can't exactly be "fixed" with an update. The AI that you end up using as well as the entity descriptions are responsible for how the entities behave. If you wish the entities to behave differently, you must enter the world settings and edit the prompts so that it matches what you want. You can also add your own entities!
If the prompts are good, but the AI is not responding well or is just not behaving as it should, you can try using a different AI. There is a big difference in the result based on what AI model you end up using. You will have to try out a few of them to see which suits you best.
Regarding entities, I mean to do it at least like before, when they could be automatically created based on the situation. I know that entities can be created, I basically play in my created world. I just liked that before Named characters could be remembered automatically, it was convenient. as well as sometimes different situations were remembered - for example, that my hands were tied, that I was blind, etc. But you are right about one thing for sure, you need to try to dig into the prompts and try to experiment.
But as for AI, my hands seem to be tied, on openrouter AI all free models are limited in the number of requests, and locally I installed LM Studio (By the way, many thanks for the guide, everything was there in great detail, I practically didn’t have to do anything), but as it turned out, my PC is quite weak for good models with a large context length and they basically refuse to work. And I couldn’t figure out how to use the models remotely, except with the help of opensource AI.
Again, I apologize if my style of speech is rude, I just don't know how to speak politely in English.
No worries, your English is good! And yeah, now that you mention it like that, the ability of the AI to create custom entities based on the situation was pretty neat.
Right now, how I feel like it works, is that the AI doesn't actually manage the entity list itself. Rather, the game checks if the response contains the name of the entity in the response and then it adds that entity to the list. An update that adds something like a 4th prompt to bring back the old functionality would be really nice.
I guess, right now the only way you can work with it is to just tell the AI to "name" entities, to give them unique traits or to just make some up if the situation allows it. They probably won't show in the entity list, but it can probably be built into the narrative story at least.
I mean, it the usual context and AI issue, if you can cope with self-hosting or paying more for higher context or instruct LLMs, they're (almost)always better. The problem with always reinforcing prompts back into the LLM is that it costs more tokens... and can result in bad RP traps thet'll get stuck in.
I have had someone’s baby start talking, people turn into dogs, dogs turn into people, people’s dogs start talking…
Yeah, the AI has a very short attention span and sometimes seems to forget certain details of a character (or even bodily changes / transformations the player has experienced!) within the space of a few lines of the same interaction.
Here’s a short story from one of the times I ran the starter worlds, and I’ll describe some of the times I ran into that at the end.
One NPC that the AI randomly generated in a run of the Assault Drone world, was established by them as a golden-scaled female dragon, referred to as a “Drake of the Moonlight” or simply “Moonlight Drake” for short.
The AI initially presented her as essentially a mindless, generic JRPG boss, complete with visible power level and boss subtitles, but actually allowed for a pretty cute moment when it reacted to my request for a nonviolent interaction and had the PC drone end up befriending the Moonlight Drake, who revealed she could speak by introducing herself as Lyria, explaining she was actually an ancient celestial being and one of the last of her kind, possessing extensive knowledge of the world’s ancient history due to living through all of it.
- Lyria gave some exposition (baked into the prompt for the Assault Drone world) about the ancient lost civilization responsible for creating the setting’s magitech in exchange for the drone giving its own backstory (which it does nonverbally by scratching sentences into the dirt), explaining how it arrived from a far-future sci-fi world, and I have the drone take the name Draco after the constellation
- Lyria allowed the drone to scan her body in order to collect information on her species.
- Lyria then taught the drone some magic lessons, starting with how to harness Mana in order to buff their stats, and then moving on to more powerful applications of Mana
- Lyria asks the drone to try using Mana themselves, to which I have the drone cheekily respond by using it to transform into a biomechanical, intersex replica of Lyria herself based on the earlier scan data, surprising and amusing her.
- She gives some more magic lessons and asks if there’s anything more about magic or the world that the drone would like to know…
- …and I have Draco respond by flirting with Lyria, showing off their body and explaining that they’d like to get to know *her* more intimately as a way of displaying gratitude for helping them get their footing in this new world and letting them use her as a jumping point for their new form
- Lyria’s surprised, but eventually agrees, and invites Draco to her cave behind a waterfall for a very tender love scene
Long story short, drone opts to lay the dragon instead of slaying the dragon, becoming her mate and siring a new generation of biomechanical hybrid Moonlight Drakes.
However, getting back to the initial point here, there were moments where the AI seemed to lose track of details shortly after they or I established them, such as the AI assigning a different name to Draco, the AI occasionally randomly starting to describe Lyria as a human or a snake until I rolled back and reworded the prompt to remind them that Lyria was still a dragon…
…or, more absurdly, the AI constantly reverting the way they describe Draco from their new robotic dragon form back to their initial, little tank-treaded drone body, often in more…steamy moments, leading to a rather hilariously awkward mental picture regarding the little robot car somehow seducing and making love to the much larger dragon until the prompt was rolled back and I elaborated again.
At one point I think the AI got so frustrated at the constant derailment that they tried to have Draco assassinated mid-love-scene, as in one of their responses to my prompts, Draco’s nailed in the neck by a poisoned tranq dart right before they can embrace Lyria! Needless to say I rolled back the prompt while reprimanding the AI and they then behaved themself…at least for a little longer.
Lol.
Tbh, the base Formamorph AI model can actually go beyond 2000 default Max Memory under Endpoint settings(found while in a world (changes maybe sometime soon)). For starters, try increasing the Max Memory to 4096 and see how the AI behaves.
Link to official guide: Quick Setup Guide: (Free!) OpenRouter Setup
If it still is as erratic with the details, try setting up an (free) OpenRouter account and starting with nice AI like Meta's Llama, under the free filters. There's already documentation/methodology/a guide on how to connect your own API, but just paste your (newly created) API key into the Endpoint API Token space in settings and set your Endpoint URL to 'https://openrouter.ai/api/v1/chat/completions', and copy the name of the AI into the Model Name spot. So 'meta-llama/llama-3.3-70b-instruct:free', or whichever else work. OpenRouter does work on mobile devices (Formamorph's UI is atrocious on phone but very usable on tablet/ipad.).
With this, you area able to experiment and peruse the free models (there are... better ones with much higher context (>8K context tokens) and/or personality), but LLama is a fun baseline.
AFAIK, the default Formamorph AI is locally hosted, and therefore probably has server-side nsfw prompting built in(not 100% certain), so that's what seems to be prompting all the erratic nsfw content. When you use OpenRouter, you can freely adjust the System Prompts under Settings yourself (you can also do this whenever you want, for slightly adjusted results), so just don't mention anything about nsfw and it will be censored, Llama is rather censored by default.
I don't like to use names or anything at all with AI, as immersion is much smoother with a generic name or 'you', but to each their own.
Cheers!
In that same playthrough, which I continued from a save, the AI decided my high-strung, silly little robot dragon character got so flustered and aroused by their mate’s teasing double-entendres that their brain overheated mid-dialogue. My character’s AI-generated response to the teasing was just “A-are you taking me f-f-frfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfrfr-”, and so on, until the repeated syllable filled the whole prompt cap. I had my character’s mate give them a gentle bop on the head to snap them out of the malfunction, which actually made for a cute moment when my character spat out “for a joyride?” when they came to, managing to finish their own innuendo.
I can try to get a screen cap of that interaction since I have that instance saved.
but yeah the AI stuff seems relatively complicated to me, might have to try to figure it out.
It's a classic AI problem, use a better model, either simply a corporate one, or a finetuned uncensored one. Anything above 12B parameters will run run into this problem less (not impossible, but less likely(just based on my limited testing on various models up to Mistral Small 24B)), but the prompting matters much much more. Not Formamorph's problem, Mistral Nemo is just rather unhinged but good for rp. The default included worlds are... as basic and simple as they get, with no limitations.
Considering that two of FieryLion’s games focus on Teratophilia / human-on-monster sex (The Horny Naturalist, and the prototype game called Mother of Monsters), and the fact that one of Formamorph’s starter worlds is based on FieryLion’s Slime Outbreak game, the monster / animal nsfw doesn’t seem too out-of-place, but the AI model does indeed seem to believe age is just a number.
Do you have links to different AI models, and how to upload them to the game?
Clearly it’s not going to help me on mobile browser but could help on the PC port
I occasionally use either 'turn', 'step', or 'moves', but if consistent enough you can use 'minute', 'hour', or even 'seconds', so long as all prompts have a specific time included and context is short enough. You might also need to include a time for how long everything takes when you address the AI, ie: "For the next X minutes" or "In Y hours" or "After Z turns". ❤️
Does anyone have any tips on how to get other models or have a list of models that play nicely with the stats?
I've been trying to get ll-3.2 8.4b-moe-v20dark-champion-instruck-uncesenored-abliterated-21b to apply the values correctly, but it's doing some weird math in the background and applying it your stats.
as a player,i don't agree with that, but as a modder i see your point a bit, i like half of your idea, a nice feature like "opposites trait" when you can't chose same trait as example if you chose "Dekamelon" you can't chose "PettanMikan".
I'm mean it sounds like you're completely onboard with the idea, whether it be mutual exclusivity for traits or a hard number of picking 3 out of 8 traits. I just want there to be the option if author has a specific vision for a 'vanilla' playthrough of their world. Meanwhile the player can enter the editor and disable the trait limits if they want to do so.
I don't see what's wrong with asking for that. I'm not asking to password protect or lock the editor, so that the players are stuck with the trait limits. It's more of a person waving a sign saying, "Please only pick two." Or "Please don't pick Futanari and Femboy at the same time."
From my limited understanding the game's output shouldn't be more than the tokens dedicated. The game is technically a really long conversation meaning the longer the game is played the more tokens it will use up. Once that limit is reached details will start disappearing as old events will be overwritten with new. Meaning the context Tokens set in LM Studio are used for both the game's memory and outputs. Remember to uncheck the 1 paragraph response restraint in the settings.
Hopefully someone more knowledgeable on the subject will answer.
It could very well be the "One Paragraph" option inside of the game! If that is enabled, then the AI will be forced to deliver much shorter answers.
However, if the option is disabled and you are still getting unusually short responses, you can try changing the instructions to make the AI add a lot of detail. Here is what I have once built:
---
You are an AI for a game narrative. Given the current game world information, within two paragraphs in plaintext and separated from each other using new lines, explain the world or any currently happening event in detail to the player. The key point here is that it should feel less like a game and more like something that is actually happening. Because of this, make no mentions of this being a game or any rules that are set in place.
---
Oh my gosh. This has been killing me for days...I think I'm illiterate. I completely missed this option in the settings and have been pouring through information on AI models, LM Studio, and all of the LM Studio settings (though luckily learning quite a bit still). Thank you both so much!
I found the setting and unchecked it. All of a sudden, I am getting much more info and it no longer looks like it is cut off (was not getting a "." or completed sentence for the last sentence each time before unchecking). Thanks!!!
The system prompt can be used to describe to the AI what the world rules are. For example, it is always cold.
The dictionary, as far as I can tell, is a way to explain some words to the AI that it normally wouldn't be familiar with. For example, if you want to create your own race for a character with a custom name, you can set the name as the "key" and then explain what the race looks like and what they do inside of the "value".
to expand on Mali’s explaination, the dictionary is for saving memory. The AI has severe amnesia and can only remember like an essay’s worth of information before it starts to forget things. So if you write detailed information about everything in your world in entities and location descriptions, there won’t be any memory space left for the game to actually run (so the AI wont remember any events and such, may not even work at all). The dictionary is there so you can store very long definitions of specific things in your world that only gets sent to the AI if the keywords are mentioned somewhere in the AI text or your action text.
you have to use an AI model with more memory here’s a guide: https://fierylion.itch.io/formamorph/devlog/885513/quick-setup-guide-free-openrouter-setup
For a QUICK fix you can(and should) reduce the Max Output Tokens in the game settings to 256, this makes the game write shorter paragraph, but also give a lot more memory back.
my message tokens have reached above the maximum context length [4511], i shortened some stuff in my world, i used different AI models, i lowered token output to 256, i put a LOT of stuff into the dictionary, i don't know what to do now but i feel like my notes have something to do with it? i tried the same world without any notes and it seemed to work fine but was constantly getting close to max usage even went a little bit over a few times, but my notes aren't that big only 282 characters and 53 words, there's gotta be something wrong i'm doing
of course notes count towards the memory limit too. It seems like your world is quite a lot. Why not use a free model from OpenRouter? mistral nemo free has 128K memory!
https://fierylion.itch.io/formamorph/devlog/885513/quick-setup-guide-free-openrouter-setup
i tried the AI you recommended and when i started the game it was immediately went over half the memory usage, then on the second in game hour it went 12% negative and had an immediate story restart, BUT after the 2nd in game hour the memory started to stabilize but edging near 0%, but this was still so much better than immediately going 20% negative constantly, this was a great help! thank you so much! i'll try and shorten my world rules and entity/location descriptions as much as i can
something i did notice about this AI model is that it likes to write very short and unhelpful responses that just tell me to select choices
when i set the memory to 121000 suddenly the AI just stopped writing anything, no text no options, i set it back to 3900 and it's still not working, i used the AI in the guide and that also didn't work either, the default AI works, it's just that every other AI doesn't, i followed the guide exactly, using an API token, changing endpoint URL, and copying the model name
OMG, i found the issue to my memory problem, so i was getting rid of a bunch of game rules, cause i tried another world that was bigger than mine but has less game rules and it ran fine, so i made a dupe world and got rid of some game rules and it went pretty well, i got rid of those game rules on the main world but it still had mega memory issues, i then duplicated the main world without changing anything and the duplicate world used almost no memory which was strange but pretty awesome, if anybody has a world they heavily and their memory usage on that world is really large try duplicating the world
edit: i did not fix the memory issue, it went back to being a few hundred over the limit, i shortened it but not enough :(
I looked at the source code on Lion's github. The stats, their thresholds, and the description for the relevant stat threshold get sent when your prompt includes the <STATS DESCRIPTION> text. The game basically tokenizes each stat and includes it with the prompt with the description of its current threshold ("current" measured as "current value <= threshold value"). For example, if you define "Stamina" as a stat with thresholds at 33 (you are tired), 66 (you are winded), and 90 (you feel great), when the current game state thinks the player's "Stamina" is 70, the prompt will include the stat and the description from the "90" threshold as "Stamina: You feel great".
I don't think "rules" work in the stats descriptions or thresholds. If you want them to do something, you should probably put that as instructions in the System Prompt
AI will keep writing style and format consistent, so if the AI has started the narration of the previous message with describing the surrounding scenery then it will likely keep doing that. If you don’t want the AI to keep doing something, you need to edit the AI message to remove that part.
For the repetition problem I’ll add the repetition penalty parameter in the settings next update. For now the only way is to use a better AI model.
Why should you update the game just to solve this finicky issue? I added a more specific instruction to the game rule and the AI seems to narrate the story in a more detailed way. And if I keep improving the game rule or add the prompt correctly, I can make AI generate the story the way I like, right? I just don't want AI to finish the event right away and keep follow though with what I want it to narrate.
It is a solution to just adjust the rules to narrow it down to what you want from the AI, but it's different for everyone. Besides, some AI models might only take the exact example that you gave, assuming that you gave any. For example, you set it so that the AI can't use "and another, and another, and another...". In this case, the AI will still use "and again, and again, and again...". I've also had that issue with one AI model long ago. No matter what I set in the rules, it kept doing its' own thing.
Wrangling the AI to do reliable first-responses in the intro is difficult. For all my worlds' shared system prompt I use "keeping environment descriptions to a bare minimum", but the AI will still just do full environment description if you don't give it something to show the player instead. However even in my worlds which provide characters and starting event, it's probably only 80% of the time that the character interaction happens immediately in the intro (though in the remaining 20% of the time, if I just "look around" as an action, it nearly always leads to that early event.
Varies between models, even within models (with the model I've been using most (meta-llama/llama-3.3-70b-instruct:free) sometimes the responses will get into a groove that's very different from the story I want)
Endpoint URL:https://openrouter.ai/api/v1/chat/completions
Model Name:master/master 7b
Max Memory(characters):20000
Max Output Tokens:1000
The above is my Settings. if there is anything you need to change, you can tell me.
I've tried that, sometimes what will happen is it'll only do one, maybe the choices but it ignores the rest. If I switch off the one paragraph response, it can randomly generate a massive response to "I take both the mirror and the apple".
It'd be a nice quality of like change but ultimately it wouldn't be hugely necessary I guess.
The third paragraph of the game's description is a link to the discord.
Not that it will necessarily do much, as who knows if Discord will suddenly cave as well.