How to use KoboldAI Lite, the chat and story UI in KoboldCpp
KoboldAI Lite is the web page KoboldCpp opens at http://localhost:5001. Pick a mode (Instruct for assistant-style questions, Chat or Adventure for roleplay, Story for co-writing), then use Memory for facts that must always stay in view, World Info for details that appear only when a keyword comes up, and Author's Note for style instructions at the end. Your work is stored in the browser, so export a save file to keep it safe.
Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
- lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
On this page
- What KoboldAI Lite is
- First visit: the welcome box
- Choose a mode
- Memory, World Info and Author’s Note
- The order the model sees
- Memory
- World Info (lorebook)
- Author’s Note
- How much context they use
- Character cards, personas and group chats
- Settings that matter
- The seven settings tabs
- Save, load and where your data is stored
- Images, voice, documents and web search in Lite
- Lite on a phone or another PC
- Lite online vs Lite in KoboldCpp
- Problems
What KoboldAI Lite is
Lite is the browser interface embedded in KoboldCpp. In 1.124.2 it is Lite version 353, and the page at / reports it. It is the same app as lite.koboldai.net; online it can connect to the AI Horde and other providers, while inside KoboldCpp it talks to your local server. KoboldCpp also serves other pages. All six answered 200 in our test F62:
/ | KoboldAI Lite |
/lcpp/ | The llama.cpp web UI |
/sdui | StableUI, for image generation |
/musicui | MusicUI |
/noscript | A page that works without JavaScript |
/api | The API documentation page |
One server answers all of these and several API families F23. See KoboldAI, Lite and KoboldCpp: what is what.
First visit: the welcome box
The first visit shows “Welcome To KoboldAI Lite” with three styles, Classic UI, Aesthetic UI and Corpo UI, and buttons to launch the llama.cpp UI, StableUI and MusicUI F31. You can change the style later in Settings.
Choose a mode
The mode list in Settings has four entries F62. Lite starts in Instruct mode F31.
| Instruct | Assistant-style questions and answers | The start mode. Pick the instruct template that matches your model: chat templates. |
| Chat | Talking with a character | Adds a Group button for several characters. |
| Story | Co-writing: you write, the model continues | The request we inspected below was made in this mode. |
| Adventure | Text adventures | You type actions. |
See chat templates. We have not captured the request shape of the other three modes.
Memory, World Info and Author’s Note
Open them from the Context button: the dialog has tabs Memory, World Info and TextDB.
The order the model sees
We inspected a real request in Story mode, with a Memory text, one World Info entry (key dragon) and an Author’s Note, and a story line that contained the key F61. Lite sent Memory and the triggered World Info text together in a separate memory field, and the story with the Author’s Note appended as the prompt:
- Memoryalways sent, first
- World Infoonly entries whose key appears; after Memory by default
- Story so faryour text and the replies
- Author's Noteadded at the end of the prompt
KoboldCpp puts the memory field before the prompt. All four count against the context size. When the context is full, the oldest story text is what gets cut: context size.
Memory
Text that is always in view. Keep it short. There is an AutoGenerate Memory button, and a “Notes” button next to it. Our request carried the Memory text first, followed by a line break F61.
World Info (lorebook)
Each entry has keys and the text to remember; the entry is added when a key appears in the recent text. In the World Info tab, you can set where entries go (“After Memory” is the default, or “Before A/N”), how far back to search (“Full Context” or the last N tokens) and whether keys are case-sensitive F61. Example for a fantasy story: key dragon, text “Dragons breathe blue fire.”
Author’s Note
Style steering. Lite adds it at the end of the prompt as [Author's note: …] F61, and offers four strengths: Weak, Medium, Strong and Immediate.
How much context they use
Every character in Memory, World Info and the Author’s Note uses context that the story can no longer use. We have not measured token counts for them yet.
Character cards, personas and group chats
Lite’s start screen links to loading a JSON file or a character card, and Chat mode has a Group button for several characters. We have not tested importing a card yet, so we cannot say which field lands where. For heavy roleplay, many users prefer SillyTavern.
Settings that matter
- Amount to generate: the server’s default generation length is 2048 F05, and the request we saw used
max_length2048. - Lite’s own context size: in our test the server was started with
--contextsize 2048, the server reported 2048, and Lite still sent a context of 16384 F61. Lower Lite’s setting to match the server. When the prompt is longer than the server’s context, the server cuts the start of it without an error F53, so the model forgets your Memory. - Instruct tags, thinking and samplers: chat templates and samplers.
The seven settings tabs
Settings (top right) has seven tabs in Lite 353, which is the version inside 1.124.2 F104: General (usage mode: Instruct, Story, Adventure or Chat; UI style; instruct tag preset; responses), GUI (themes, colour schemes, markdown, extra pages), Samplers (samplers), Media (images and voice), Tokens (stop sequences, logit biases, phrase bans, thinking tags), Tools (MCP and web search, embeddings) and Misc (autosave, advanced options, reset). The Context button opens Memory, World Info and TextDB.
Save, load and where your data is stored
| Autosave of the story and settings | Your browser: an IndexedDB database named e_kaihordewebui_, plus two small markers in localStorage | F62 |
| Which address | Each browser origin has its own storage: localhost:5001, 127.0.0.1:5001 and a tunnel link do not share saves | browser rule, not tested |
| An exported save file | A file you choose, from Save / Load | upstream UI |
| Server-side saves | A file on the server when KoboldCpp starts with --savedatafile | upstream flag |
| A preloaded story | Loaded at start with --preloadstory | upstream flag |
| Generated images | Not checked | not tested |
Users report that Lite slows down with very large saves (#44, #492): archive old text and summarise it into Memory.
Images, voice, documents and web search in Lite
- The Add File button opens “Add Audio or Image”: Generate Image (Automatic or From Prompt), Generate Speech, Capture From camera or microphone, Upload a File (image or audio), paste, and a shortcut to the media settings F105.
- Images: image generation.
- Voice: voice. Users report that the microphone needs HTTPS or localhost (#920).
- Vision: vision.
- TextDB and web search: embeddings.
Lite on a phone or another PC
Reach it over your Wi-Fi or a tunnel: remote access.
Lite online vs Lite in KoboldCpp
| lite.koboldai.net | Lite inside KoboldCpp | |
|---|---|---|
| What it is | A hosted page | The copy served by your own server |
| Who generates | The AI Horde or a provider you choose | Your own model |
| Where the prompt goes | To that provider | Only to your computer |
See AI Horde.
Problems
| What you see | Cause | Page |
|---|---|---|
Failed to fetch or NetworkError | Lite lost the server | connection problems |
| Nothing happens on send | The welcome box is still open (F31) | welcome box |
| Replies are cut short | The amount to generate, or a normal end of reply | samplers |
| Slow with long stories | A large save or a re-read prompt | speed |
Frequently asked questions
What's the difference between Memory and World Info?
Memory is text that is always sent. World Info entries are sent only when one of their keys appears in the recent text. In our test, both went into the same memory field of the request: Memory first, then the triggered entry.
Where does KoboldCpp save my chats and stories?
Lite autosaves in your browser, in an IndexedDB database. It is not a file on disk, and it belongs to the address you opened (localhost, 127.0.0.1 and a tunnel link are separate). Export a save file to keep it safe, or start KoboldCpp with --savedatafile to keep saves on the server.
How do I import a character card into KoboldAI Lite?
Lite's start screen has a link to load a JSON file or a character card. We have not yet tested importing a card, so check what lands in Memory after you load one.
Is KoboldAI Lite private?
The copy inside KoboldCpp talks only to your own server, so your text stays on your computer. The hosted lite.koboldai.net page can send prompts to the AI Horde or another provider you choose.
Which mode should I use for roleplay?
Chat or Adventure mode for roleplay, Story mode for co-writing, and Instruct mode (the start mode) for assistant-style questions. Pick the one that matches how you want to type.
Why does Lite reprocess the whole story after I edit Memory?
Users report it. Memory is sent before the story, so changing it changes the start of the prompt, and the server has to read the prompt again from that point. We have not measured the cost.