ReAct Loop

The ReAct Loop is the minimal execution unit in Memfit AI, driving the "Reasoning-Acting" cycle within an independent context.

Core Components

LoopEngine

The LoopEngine manages the state flow of each ReAct iteration:

• Handles state transitions
• Manages iteration limits
• Tracks execution history
• Coordinates with reliability mechanisms

Action Registry

The Action Registry maintains available actions:

• Registers custom actions
• Maps action names to implementations
• Validates action parameters
• Handles action lifecycle

Prompt Manager

The Prompt Manager generates context for LLM:

• Constructs system prompts
• Injects tool descriptions
• Adds memory context
• Formats conversation history

Execution Cycle

Each ReAct Loop follows this cycle:

┌──────────────────────────────────────────┐
│              OBSERVATION                  │
│   Collect environment state & feedback    │
└─────────────────────┬────────────────────┘
                      ↓
┌──────────────────────────────────────────┐
│               THOUGHT                     │
│      LLM reasoning & spin detection       │
└─────────────────────┬────────────────────┘
                      ↓
┌──────────────────────────────────────────┐
│               ACTION                      │
│    Execute tool or request information    │
└─────────────────────┬────────────────────┘
                      ↓
                 (Next Iteration)

Focus Mode

Focus Mode allows defining specialized execution contexts:

Use Cases

• Code Audit Mode - Specialized for code analysis
• Vulnerability Scan Mode - Optimized for security scanning
• Report Generation Mode - Focused on documentation

Implementation

A Loop can be encapsulated as an Action, allowing:

• Push current context onto stack
• Start a specialized sub-Loop
• Complete focused task
• Pop and return to parent context

Dynamic Capabilities

Suspension & Escalation

The ReAct Loop supports:

• Dynamic Suspension - Pause for human input
• Help Requests - Ask user for clarification
• Plan Triggering - Escalate to Plan Engine for complex subtasks

Mode Switching

Seamless switching between:

• Instant execution mode
• Plan-execute mode
• Focus mode

Reliability Features

Built-in Protection

• Self-Reflection - Analyzes failures within the loop
• Spin Detector - Prevents infinite loops
• Step Limits - Enforces maximum iterations

Error Recovery

When errors occur:

1. Capture error context
2. Trigger self-reflection
3. Generate correction plan
4. Retry with enhanced awareness

Configuration Options

Option	Description	Default
maxSteps	Maximum iterations	20
spinThreshold	Steps before spin detection	3
reflectionTrigger	When to trigger reflection	On error
focusMode	Specialized mode name	None