Feature Cognitive Routing

Key Feature: Cognitive Routing & The Zero-Dollar Brain

The Problem: Running a background memory refinement pipeline (summarizing 100,000 tokens of raw terminal logs every hour) using a flagship model like Gemini 1.5 Pro or GPT-4o will rapidly drain your API credits.

The Solution: Cognitive Routing. A.I.M. splits the "Brain" into discrete, interchangeable LLM endpoints, allowing you to use hyper-expensive models for logic, and 100% free local models for memory management.

---

The Universal Hub (The Cockpit)

A.I.M. is completely model-agnostic. The aim tui command opens the Sovereign Cockpit, allowing you to route different tiers of the architecture to completely different API providers (Google, OpenAI, Anthropic, OpenRouter, or Local Ollama).

1. The Frontal Lobe (Maximum Reasoning)

• Assignment: The Default Agent.
• Model: Gemini 3.1 Pro or Claude 3.5 Sonnet.
• Use Case: Writing complex Rust backends, debugging React states, and making deep architectural decisions. You pay the premium API price because you need maximum logic.

2. The Muscle (The Zero-Dollar Brain)

• Assignment: The Tier 1 Harvester / The Indexer.
• Model: Local Ollama (Llama-3 8B) or Gemini Flash.
• Use Case: Reading 10,000 lines of JSON to extract the "Signal Skeleton."
• The Secret: Smaller models typically fail at complex RAG pipelines due to hallucination or sequential tool spam. A.I.M. solves this via the Explicit Guardrails toggle during aim init.

The "Token Tax" (Explicit Guardrails)

When you assign a local, lightweight model to handle your background tasks, A.I.M. automatically injects a rigid, ALL-CAPS behavioral matrix into the prompt:

## ⚠️ EXPLICIT GUARDRAILS (Lightweight Mode Active)
1. **NO TITLE HALLUCINATION:** You MUST NOT guess file contents from titles.
2. **PARALLEL TOOLS:** You MUST request all files in a single tool turn.
3. **DESTRUCTIVE MEMORY:** You MUST delete stale facts.

By imposing this strict "Token Tax" on the prompt, A.I.M. physically forces low-IQ local models to behave like disciplined data-entry clerks.

The Result

You experience the brilliant coding capabilities of a flagship model in your terminal, while your massive background memory cascade runs entirely offline on your GPU for $0.00.