Model routing
The auto model: ladders, the classifier and the agent-mode floor.
Ask for model auto and Memtro picks a rung on a ladder of models, cheapest to strongest, based on how hard the latest message is.
How a decision is made
- Heuristics score the request from its shape: greetings and acknowledgements are trivial; short plain questions are simple; code, multiple questions, long conversations, images and "design the architecture" style words push the level up. Clear cases are decided instantly.
- Everything else gets one short classification call on a cheap model (the cheapest rung, or an API-backed cheap model when the cheapest rung runs through the Claude Code CLI). The decision is cached for fifteen minutes.
- The level (1 to 5) maps onto the ladder; clamps from your settings apply.
- Agent mode never routes below level 3, so a capable model drives multi-step investigations even for a one-line request.
Responses carry the decision in memtro.routing (complexity, model, method, reason), and the Chat page shows it under each reply.
Default ladders
| Provider | Ladder (cheapest → strongest) |
|---|---|
| Anthropic | Haiku 4.5 → Sonnet 5 → Opus 5 → Fable 5.1 |
| OpenAI | gpt-5.4-nano → gpt-5.4-mini → gpt-5.4 → gpt-5.5 |
| Gemini 3.5 flash-lite → 3.8 flash → 3.1 pro | |
| Claude Code | haiku → sonnet → opus → fable (aliases the CLI resolves) |
auto/<provider> stays within one provider. Plain auto uses the configured ladder, or the default ladder of the first provider you have a key for.
Settings
Dashboard → Routing lets an organisation set the ladder (one model id per line), the classifier model, minimum and maximum levels, and whether the memtro default alias should be routed too. Each user can override these for their own requests. The page includes a tester that shows how a message would be classified.
The same page holds agent mode settings: the default mode, the step budget, the delegate model, whether to stream progress, and which engine runs agent mode.