Chat Optimization
Dashboard settings for how long the AI's replies are, how much of the conversation it remembers, and a fix for models that need the conversation to end with your message. Open the panel at Chat Optimization in your dashboard.
AI Response Settings
Pushes models toward longer, more detailed replies. Pick one, or leave it at None.
| Value | Mode |
|---|---|
| 0 | None (default) |
| 1 | Force medium responses |
| 2 | Force medium responses, roleplay style |
| 3 | Force medium responses, roleplay style with inner thoughts |
| 4 | Force full responses |
| 5 | Force full responses, roleplay style |
| 6 | Force full responses, roleplay style with inner thoughts |
User Role Injection
Some models refuse a conversation that does not end with a message from you, for example one that ends with an unfinished AI reply (a prefill). The request then fails with an error such as Requests ending with a model turn are not supported, This model does not support assistant message prefill or Last message cannot be from the assistant.
Turn on Enable User Role Injection in Chat Optimization and LiteRouter adds a short, neutral placeholder message with the user role at the end, so the request is accepted:
- Claude and Gemini models: whenever the conversation does not end with your message.
- All models: when the conversation has only system and AI messages and none from you.
Nothing is added when the conversation ends with a tool result, because the model can already reply to it. Your messages are sent exactly as they are, and the placeholder carries no instructions. The setting is off by default; when it is off, the conversation is sent as it is.
Chat memory
Two multipliers set your context window. Each 1× is worth a fixed number of tokens:
| Component | Holds | 1× is worth |
|---|---|---|
| System Context | The rules and role the AI is given, sent with every request | 1,250 tokens |
| Conversation Context | The 5 latest messages, 750 tokens each | 3,750 tokens |
Set both sliders to the same number and it multiplies out: 1× gives you 5,000 tokens, 5× gives 25,000, and 10× gives 50,000. Ordinary :free variants stay at 5,000 on every plan. Listed :full-context and named context variants such as :32k-context use their own context settings instead; see Model naming.
Your plan's ceiling
The sliders move independently. Your plan caps a weighted score of the two, with conversation counted three times as heavily as system:
((1 × System multiplier) + (3 × Conversation multiplier)) / 4Say your plan allows 25×, and you set System to 20× and Conversation to 24×:
- Score: (20 + 3 × 24) / 4 = 23. That is under 25, so the plan allows it.
- Window: 20 × 1,250 = 25,000 tokens of system, plus 24 × 3,750 = 90,000 tokens of conversation, so 115,000 in total.
Claude models have separate chat memory settings below the general ones, adjustable from the Plus plan up. Their managed-context ceiling is 5,000 tokens on Basic and Standard, 25,000 on Plus, and 50,000 on Pro, Elite, and Ultimate. Higher plans have larger ceilings. Model access depends on the minimum plan shown in Models; a context ceiling does not grant access to every Claude model.
What it costs
For ordinary premium variants, longer requests spend more credits:
- The first 15,000 tokens of a request count as 1.
- Every 5,000 tokens after that adds 1.
- Multiply the total by the model's cost. The cost calculator does this for you.
Cost climbs as a chat gets longer, until it reaches the ceiling your multipliers allow. Output tokens are not counted, though the reply joins the conversation on the next request.
Choosing your settings
Start at System 3× and Conversation 2×, then raise what you need.
Some frontends, SillyTavern among them, send character and user personas as conversation messages instead of system ones. On those, keep both multipliers at 5× or higher, or the personas fall out of the window and the AI forgets them.
Individual messages are never trimmed. If one runs past 750 tokens, fewer messages are kept.
Reset to Default returns every multiplier to 1× and the response mode to None.