Starter Guide

On first launch after installing, you'll see the onboarding wizard — just follow its prompts step by step. It will have you link a few basic settings, and then you're in the app.

To chat at all, you first need a "chat model" — it's a prerequisite.

After onboarding you'll usually land on the "Me" tab. If you already configured a chat API in step ①, skip ahead; otherwise continue:

Not sure what these are or where they come from? Search or ask an AI. Beginner keywords: "What is an AI API? Where do I get one? How do I use it?"

Tap the "Settings" gear at the top-right → "API Setup" → "Model Manager" → the "＋" at the bottom-right, then set it up:

1. Config name: Just a label for yourself — not the model ID!
2. API type: If unsure, pick the first one, "OpenAI Compatible".
3. API endpoint: The endpoint URL your model provider gives you. They usually provide two (OpenAI Compatible / Anthropic Compatible) — match it to the type you chose.
4. API key: The secret key you created in the provider's open platform.
5. Model name: That's the model ID, listed by the platform, e.g. deepseek-v4-pro. Many providers' model IDs are strictly case-sensitive.
6. Max tokens: The length cap of a single reply. The default 4096 is plenty for daily use; raise it only for very long code or articles, otherwise keep the default.
7. Context window: How much it can keep in mind at once. The value must be ≤ what your model supports (if it maxes at 200k, don't enter 1M). Within the supported range, bigger means it remembers more. Leaving it blank / 0 usually falls back to the model default.
8. Vision (image input): Whether the model natively supports image recognition. If unsure, search or ask an AI — don't enable it if unsupported.
9. Thinking mode: Follow the default; only change it if a provider's model ships with thinking off.
10. Temperature / top_p and similar: You basically won't need these for normal chat. If you have special needs and don't know where to start, the app usually has a brief explanation, so we skip it here.
11. Test connection: Tap it and wait for the result. Success means it's correct. On failure, check the model name / key / endpoint, make sure there are no wrong symbols or trailing spaces, that your platform balance isn't empty, and it's not a peak-hour rate limit. If the error is confusing, screenshot it and ask an AI. Finally, don't forget to tap "Save config".

To make the AI's memory "smarter" (recalling related old memories by meaning), you can set up an Embedding model if you can — it's not required.

Not sure what it is or where to get it? Search or ask an AI. Keywords: "What is an AI vector-search API? Where do I get one? How do I use it?"

From ②'s "API Setup" → "Vector Search" → the "＋" at the bottom-right, then set it up:

1. Config name: Same as ② item 1.
2. API endpoint: Same as ② item 3.
3. API key: Same as ② item 4.
4. Embedding model name: Same as ② item 5.
5. Test connection: Same as ② item 11.

To let the AI search live web info, you can set up a web-search service if you can — it's not required. Web search is a third-party service, so you configure it yourself.

Not sure what it is or where to get it? Search or ask an AI. Keywords: "What is an AI web-search API? Where do I get one? How do I use it?"

From ②'s "API Setup" → "Web Search" → the "＋" at the bottom-right, then set it up:

1. Enable web search: A master toggle.
2. Search service: Three choices (Tavily / SearXNG / custom endpoint). Configured much like the models; check the relevant provider's site for details.

"Me" tab → "Settings" gear → "App Settings". A few entries worth caring about:

1. Language: The UI display language — literal meaning.
2. Frame rate: Still in Beta; picking "Auto" is enough.
3. Sticker pack manager: Add and manage your own stickers here. Then, in a chat, enable the "Agent" and "Sticker" tools and you'll see them in the sticker button on the right side of the input box to pick and send.

"Me" tab → "Settings" gear → "Chat Appearance". Three personalization options (chat background, message-bubble translucency, bracket style) — tweak them however you like.

If you already set this up in ①, you can skip this. To receive proactive messages in time and keep the memory system running reliably, you should set these three:

1. Battery optimization: Set to "Unrestricted" — prevent the system's battery optimizer from killing the background process.
2. Auto-start management: So it can relaunch itself if the system kills it.
3. Persistent status-bar notification: Keeps a permanent notification so you know the app is still running.

This step is required — it's like adding a "contact". Only with a TA can you start a chat.

Tap "TA" in the bottom nav → the "＋" at the top-right → enter the creation page:

1. Nickname: Just a label for yourself; the AI won't know it. Only you see and use it.
2. Name: Unlike the nickname, this is TA's real name. You can edit it later in the persona file.
3. Persona description: This is TA's persona. Don't know where to start? Grab a ready-made template from the community (Personas) — you can always change it later, so no worries.
4. Model choice: Pick the "chat model" you added in ②.
5. Custom conversation ID: Generally ignore it.
6. After creating, tap "Done" and the contact shows up in the TA section.

Tap the contact you created in ⑧ to enter its chat. Step out, then tap "Messages" and it appears in the message list; long-press or swipe to pin it or not. Tap back into the conversation to enter the chat page:

1. Token ring: If you set a context window in ② item 7, there's a ring at the top-right showing a number — this is the token ring, for viewing detailed context usage, cache hit rate, etc. (not 100% precise).
2. The "＋" panel left of the input box — things inside it can be tapped or long-pressed:

a · Model: Switch between the chat models you configured in ②.
b · Chat settings: Three Thinking-display modes to taste; token-usage display and AI message streaming also to taste. "Show input status": while the AI types, it shows a custom or default text at the top-left. At the bottom, time awareness has two flavors — Smart mode (saves a little token on models like DeepSeek; pick this for daily use); if you need very strict time accuracy, choose "inject every message".
c · Agent (tools): The built-in tools in the app. When on, the built-in tools are available; when off, they aren't. The "tool loop protection" is fine at its default of 5 rounds; on pages two and three each tool can also be toggled individually.
d · Tool status: If the AI uses a tool while chatting, a tool bubble appears; turn it off to hide whether the AI used one. The bubble has "mini / full" styles — choose whether and how to show it.
e · Scheduled tasks: Lets the AI message you or remind you of something after a set time. With time-awareness guidance, if you say something like "I'll take a shower, let's watch TV after", a smart AI may create a reminder about 30 minutes later — it figures you'll be done showering by then.
f · Web search & web fetch: The AI's web capability in chat. If you configured ④, "web search" works; even if not, "web fetch" still works.
g · Usage habits: Your phone app-usage duration. The "desensitization level" has full / category modes — set to taste; "app scope" too.
h · Diary access: If you enabled the diary feature and the "diary access" tool, you can't directly read the AI's diary — you must ask the AI itself; it can decide how many minutes to show and which entry.

Tap the AI's avatar in the chat page, or the three-line icon at the top-right, to enter AI's home. Here you can set the AI's nickname, bio signature, and related options. Tap the "Settings" at the top-right to open the settings page — three sections matter here:

a · Message settings

a1 · Proactive messages:
· Frequency & time slots: After enabling proactive messages, the AI sends messages per your rules. The [cooldown] is how long after your last message it waits before sending one, so it won't interrupt mid-conversation; [daily cap] is how many proactive messages it can send per day; [time slots] set when it's allowed to trigger; [trigger probability] and [min / max interval] are to taste — each has a fairly detailed explanation.

· Template & inner OS: A highly customizable template; the default is good enough. [Show OS] adds an extra "OS" bubble with proactive messages; [usage-habit injection] adds "your phone app usage today" into the message — it needs the "usage habits" tool enabled first.
· Rhythm & emotion: For richer emotional color in chats, enable both "interval fluctuation factor" and "emotion system". Among them, [sense of proportion] stops proactive messages from feeling impersonal; turn it on for messages that feel more grounded and measured, according to your taste.

a2 · Segmented sending: When on, the AI's message splits into multiple segments. The app can't force it, but most of the time the AI follows — it depends on the model. If it keeps failing, turn on "long-message fallback splitting" to force segmentation.

a3 · Typing delay: Simulates the lag of a human typing. There's a "text-adaptive" mode, or a fixed per-segment delay — set to taste.

a4 · Spark: Everyone knows this one; no need to elaborate.

b · Memory system

When the memory system is on, the AI generates memory files; the "memory star map" also shows a starfield made of memories.

b1 · Context trimming: There's a strict-ish default (max context turns / max context tokens). If you care about long, stable memory instead of dropping bits as you chat, set it to "Unlimited" and let the context length of the model you configured in ② decide.

b2 · Daily memory flush: When on, it also generates memory files and consolidates memory. [Check interval]: the longer, the less often it consolidates and flushes. [Daily batch interval]: keep the default. [Flush on trim]: when a chat hits the length limit an extra trim triggers — this toggle controls whether to also flush then; keep it on so memory is lost as little as possible.

b3 · DeepDream distillation: Recommended for long-term memory (currently off by default). [Lookback days]: use the default. [Max record]: if you want memory that's more detailed and complete, raise the word count — to taste. [Include proactive messages in consolidation]: whether the AI's proactive messages are consolidated — recommended on.

b4 · Custom prompts: Highly customizable prompt templates; the default is usually enough.

b5 · AI diary: Whether the AI is allowed to write a diary.

b6 · Retrieval mode: Includes FTS5 keyword search (no vector model needed); the vector retrieval below needs the vector model from ③. The parameters below are fine at their defaults. [Index chat history] is recommended on for a stronger memory system, but it costs more tokens.

c · Skill management

Generally you can leave this alone.

From AI's home, go into AI's circle page → "Settings" at the top-right → circle settings. After enabling "circle background interaction", you can configure it.

With everything set, you can enjoy a near-fully-configured chat — send TA a message and give it a try! The rest are features you'll understand at a glance and that barely matter, left for you to explore Solme at your own pace.