Join Discord

Yuralume

AI characters
that live
between chats.

Most AI characters disappear when the chat closes. Yuralume gives each character a day of their own: memory, schedules, proactive messages, and a LumeGram feed that continues between conversations. It is self-host first, so your relationship data stays on your machine by default.

Self-host in One Line Try the Sample Chat Want it hosted?

One-line self-host is live · BYOK, your data stays on your machine.

The core runs on your own machine today; hosted Cloud and Marketplace are on the roadmap. I'm looking for early players who'll report rough edges and help smooth it out.

memory + schedules proactive Telegram / LINE path local data by default built solo in Taiwan

Demo

See how characters live between chats.

Try a lightweight simulated scene first (pre-written, not live AI), then watch the current alpha clips for creation, schedules, proactive outreach, and LumeGram memory.

Tier 0

While you're away, Mira is still living her own day.

No account, no memory writes, no background loop. This is a pre-written simulation, not live AI — it shows the first impression the hosted demo is meant to create. The live version below generates this in real time.

Before you try it live

The live demo runs on a single GPU workstation in my home in Taiwan — not a data center. It's small on purpose, and I'd rather be upfront about it:

A few live seats at a time. If they're full you'll be invited to the waitlist instead of hitting an error — it's busy, not broken.
Runs on my own GPU. Her chat and her background "day" both use local open-source models, so replies can queue when the machine is busy.
Demo caps: one character, ~80 messages per session, ~1 AI image a day; voice and video are off.
Characters are temporary. A demo character and her memory are kept ~3 days, then recycled to free a seat for someone else.
Want it uncapped? Self-host it or bring your own API key — then none of these limits apply.

Try Live with Discord Try Live with Google

Character creation flow

Watch the current alpha flow for creating a character and preparing the first self-host experience.

Schedules, story events, and Telegram

Watch a character move through daily plans, main story beats, event gacha prompts, and proactive messages that continue through Telegram.

LumeGram and memoir demo showing a character memory screen and conversation

LumeGram and memoirs

See how a character's social feed, memoir highlights, and ongoing conversation share the same relationship context.

Six designs

What makes a character feel like a person.

Not a rule-scripted reply box, but a character who remembers, has a day, and cares about context.

◐

Schedule-driven presence

Each character has LLM-planned daily activities. Completed activities leave emotional residue and can become memory.

⌖

Five-layer relationship model

Identity, life, emotional, interaction, and trust layers are built per character. New characters do not inherit old relationships.

↯

Quiet proactive messaging

A heuristic gate, LLM intention judge, and LLM decider all need to agree before a character interrupts you.

⇋

One memory across channels

Web, Telegram, LINE, Discord, and WhatsApp share the same memory, state, and schedule pool. Different interfaces, one relationship.

◬

API-first media and voice

Image, video, and TTS are API integrations kept outside the one-click core. Use BYOK providers or point Yuralume at your own ComfyUI/TTS endpoints.

✶

Vulnerable-data protection

Personal vulnerabilities can only be remembered protectively. They must not become leverage, punchlines, or manipulation fuel.

Founder Note

Built by one developer, from Taiwan.

I am building Yuralume because I want AI characters to feel less like reply boxes and more like people with continuity: they remember, have plans, send messages only when there is a reason, and keep a life outside the chat window.

The current alpha is intentionally practical. I am using lean infrastructure and a hybrid setup to validate demand before committing to heavier cloud costs. Expensive media and model-routing features will be productized only after testers show they are genuinely useful.

What is real now

Character creation, schedule and story loops, proactive Telegram outreach, and LumeGram memoir demos are already recorded from the alpha.

Who I need first

Self-host users, local LLM builders, roleplay testers, and creators who can report where setup or character continuity feels wrong.

The promise

The self-host core should stay complete. Cloud services should sell convenience, routing, support, and creator infrastructure instead of locking away character data.

How it works

Memory, schedules, and proactive triggers.

Characters do not only answer when asked. They carry context forward and act on it conservatively.

Memory model

Each turn extracts durable memories and stores them by relationship layer. Retrieval is semantic and situation-aware, not plain text search.

identity / life / emotional
interaction / trust
per-character isolation
local pgvector storage

Daily schedule

The character has a daily plan shaped by personality, world context, recent events, and story arcs. Finished activities influence tone.

LLM-planned days
activity aftermath
world fact injection
multi-day narrative arcs

Proactive loop

A scheduler evaluates whether there is something worth saying. Most ticks stay quiet; only meaningful reasons reach the user.

heuristic + intention + decider
defer / busy-reply follow-up
automatic feed posts
cross-channel fan-out to chat apps

Self-host first

The core experience runs on your machine.

Yuralume starts with people who are willing to self-host. You choose the model stack, policies, and integrations; Cloud features will be optional convenience layers, not a forced gateway for character data.

Self-host · Available now

Self-host Core

Self-host is live today: one-click deployment, character creation, long-term memory, schedules, proactive messages, and provider configuration are all in. Self-host users can bring compatible community models, including permissive, uncensored, or NSFW roleplay models where legal and intentionally configured.

Chat / memory / persona / story core
BYOK LLM and embedding providers
Compatible local, API, and community LLMs
User-controlled model and moderation policy
Image / video / TTS via external APIs
Custom ComfyUI or TTS server endpoints
Lightweight core with provider configuration
Web / Telegram / LINE / Discord / WhatsApp continuity
LumeGram character social feed
Free for personal self-host use · commercial hosting needs a license

Cloud · Coming Soon

Optional Cloud Services

Cloud will start from real self-host pain: webhook relay, tuned prompts, setup support, and later hosted or managed self-host options. Future Cloud services will also explore curated community model variants for different roleplay styles.

Webhook Relay · planned
Tuned Prompts · planned
Community model variants · planned
Roleplay and mature-audience model options · planned
Setup / onboarding support · early validation
Hosted Cloud · later
Creator Marketplace · later
Self-host messaging adapters remain first-class

Self-host · available now

Run it yourself, in one line.

One command brings up the whole stack with Docker. Your characters, memories, and chats stay on your machine. Add your own LLM key after install — nothing is metered, nothing phones home.

macOS / Linux

curl -fsSL https://yuralume.com/install.sh | bash

Windows · PowerShell

irm https://yuralume.com/install.ps1 | iex

Watch it install in 3 minutes

No speed-ups, no edits — the whole thing, from one command to a character you can chat with.

Install Docker, then run the line above.

Open http://127.0.0.1:8012 → Admin → Provider Keys and add one LLM key.

Create a character and start chatting. Full guide →

Use cases

Characters are more than chat.

Right now I'm validating companionship, interactive fiction, character feeds, and creator workflows — they shape what's worth expanding next.

Use case 1

Personal companionship

Long-running memory and schedules let a character remember yesterday and continue living while you are away.

Use case 2

Virtual relationships / fiction

Multi-day narrative arcs and branching encounters make the experience participatory instead of purely scripted.

Use case 3

Creator workflows

Character cards, stories, worlds, and branching content will start as showcase and alpha feedback before Marketplace work.

Use case 4

External message channels

Web, Telegram, LINE, Discord, and WhatsApp can share memory, schedules, and proactive messages, so a relationship continues where the user actually talks.

Two ways in

Run it now, or wait for hosted.

Self-host is live today (the one line above). If you'd rather not run Docker yourself, leave your email for the managed version — and either way, Discord is where the real conversation happens.

Hosted Cloud · coming

Don't want to self-host?

Leave your email and I'll send one message — only when hosted Yuralume is ready: I run it, your characters and memories live on, no terminal required.

Managed for you — nothing to maintain
No Docker, no terminal
I'll share Marketplace progress too

Want to be a hands-on tester? Tell me about your setup (optional) →

Community

Installed it? Say hi

Self-host phones home for nothing, so Discord is the only way I know you exist — and where you get setup help, devlogs, and demo updates.

Development updates and demos
Self-host setup support
Character card and workflow discussion