itch.io

There's an example included in the folder you can open from the settings menu. Just copy that folder, replace the assets inside, and change the options in the CharacterConfig.json file.

I'd love to when it's done, but steam currently has a special policy for live-generated AI content:

Live-Generated: Any kind of content created with the help of AI tools while the game is running. In addition to following the same rules as Pre-Generated AI content, this comes with an additional requirement: in the Content Survey, you'll need to tell us what kind of guardrails you're putting on your AI to ensure it's not generating illegal content.

The only exception to this will be Adult Only Sexual Content that is created with Live-Generated AI - we are unable to release that type of content right now.

Meaning as of now I wouldn't be able to release it since adding guardrails would go against the spirit of the game.

Now that saves are cross-version there is no reason to keep it since the purpose of that feature was to be able to keep the NPCs memories when having to start a new save for a new version of the game.
Is there another reason you'd like to have that feature for?

It works for me. Are you using high enough values? Did you damage the JSON structure (e.g. removed commas)?

You have to suspend your disbelief for this game because current LLMs are sort of incapable of admitting that they don't know something.

You can't yet.

You can send me the offending save file at HeyHeyMonika123@protonmail.com and I'll take a look (I won't look at your conversations).

There isn't really anything that should be able to pile up, and if there were it would inflate the size of the save file.

Do NPC's base your appearance off of the portrait, or can you write in your own description somewhere? Is it possible to upload your own portraits?

They're based off of descriptions. I've been thinking about having a system for adding custom player characters, so I might add that with the next update.

Especially Aldric, he'll often make up items, say he's put them in my inventory

That's called a "hallucination" and it's an unsolved problem with current LLM technology. Unless the NPC is opening the trade menu for you or straight up gifting you an item, which they can do (but sometimes fail to do), they're hallucinating and making things up.

I'll think of something.

Does it still happen too often in normal dialogs? I could remove the function from just multi-character dialogs instead.

I thought about it, but it's too much work without standardized 3D models.

After ending a dialog with an NPC, you can continue it by talking to them again within 10 turns as long as you don't leave the room (which also allows you to do things like unequipping your weapon during dialogs), so you can save and load within those 10 turns, but you have to be quick if realtime mode is active.

That's the bulk/weight bar. The changing part is a bug.

It lets you assign hotkeys to usable items. Right click to assign, and left click to use. You can also use the number keys 1-5 to use them.

I've never encountered this but I'll look into it.

It just skips a single turn, or half a minute of game time.
The game is turn based like a dungeon crawler.

#1 Is a problem with the model itself. The AI gets passed information like when the character went to bed or got up, Nemo (the model) just doesn't have the brains to connect the dots sometimes. Nemo is a 12 billion parameter model so that's one factor, but even 70-400 billion parameter proprietary models by OpenAI or Anthropic blurt out nonsense sometimes.

What is needed is a proper reasoning model alla OpenAI's o1 or DeepSeek's r1 that thinks about the situation for a dynamic amount of time before talking. It will be a while before we see an open model like that in the same size class as Nemo though.

#2 I'm trying to find ways to curb this but again, Nemo is just very likely to nonsensically order the NPC to follow you. I've already limited certain function calls to certain keywords in the dialog, but once Nemo gets the opportunity to call the follow function, it will do so for reasons I don't understand. I should have the NPC eventually give up though.
The AI that does the function calling also only "exists" for the moment of the decision after the NPC says something (that's the little "thinking" label on the input box). Having the same AI do both the talking and the function calling is too confusing for Nemo. A reasoning model would be able to handle this in one go.



I've also had it where I should've been speaking to two people at once

This should be possible with Nemo, but it's not implemented yet.

And a way to pass longer stretches of time.

Check the abilities tab. I will make this more obvious though.

I'm not sure yet. I don't want there to be too much vertical progression.

I built this version with IL2CPP for better performance, which caused an issue with starting the process.
Thanks for reporting the problem.

I limited it to about 33 words as a lazy way to have guarantees on the maximum possible amounts of tokens left after updating the long term memory of an NPC without having to do multiple API calls asking for tokenization, since I didn't think people would type that much in one message.

I'll find a way to uncap it again for 0.4, which will probably be ready on monday.

I'm at a loss how this bug made it through the cracks since I sold a lot of items during testing but I'll release a patch for it tomorrow.

I double-checked everything on my end, so my only guess at this point is that's your're using an ancient CPU without AVX2 support, or something is preventing KoboldCPP from creating the _MEI folder in temp.

If you're using a very old CPU (pre 2013 for intel and pre 2015 for AMD) you could try replacing your koboldcpp.exe with https://github.com/LostRuins/koboldcpp/releases/download/v1.76/koboldcpp_oldcpu.... and/or selecting "Use Vulkan (Old CPU)".
If you're not using an ancient CPU you could try disabling your antivirus, but I doubt that's the reason.

This seems very exotic so I can't look into it until I come back home, but it should work if you select Vulkan instead of cuBLAS.

If you only have 6 gb of graphics memory combined with 16 gb of ram, it might choke while loading the model if you have your other applications taking up too much ram.

If you have exactly 10 gb of graphics memory, the game will fully offload the model to the GPU which should take 9.17 gb, but I have no way of verifying if this actually works under real-world conditions since I only have 8 gb myself.

To see what the problem is, open a command prompt in the StreamingAssets folder like on the picture and do this:

koboldcpp.exe --model "model.gguf" --usecublas --gpulayers 17 --multiuser --skiplauncher --highpriority

Replace --usecublas with --usevulkan if you don't have an Nvidia GPU.
Replace 17 with 27 if you have 8 gb or vram, or 43 if you have 10 gb or more.

This way it won't close when it throws the error.
If it's running out of memory it should throw something like:

ggml_vulkan: Device memory allocation of size 1442316288 failed.
ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
llama_kv_cache_init: failed to allocate buffer for kv cache
llama_new_context_with_model: llama_kv_cache_init() failed for self-attention cache

You can use the KoboldCPP GUI to manually adjust things until it works if this is the case, since the game just use whatever KoboldCPP process is running in the background if you launch the game while KoboldCPP is already running.

Do let me know what the problem is though so I can fix it for the next version.

Yes.

It's just the vanilla instruct version of Mistral AI's NeMo 12B (compressed by bartowski). Nothing you type ever leaves your machine, dialog is only stored in your save file.

Everything should set itself up automatically when you start the game, as long as the model file is in the right place.

If you don't have the game on an SSD it might take a while before it's done loading.