1. Wrap neuri tool invocations to catch ActionError and return as result strings instead of throwing. This allows the LLM to learn from tool failures during its reasoning phase. 2. Enhance craftRecipe to use the recipe planner for auto-crafting intermediate materials. If the bot has raw materials (e.g., logs) but not direct ingredients (e.g., planks), it will automatically craft the intermediates first. 3. Add memory note about the dual tool interface discovery (neuri native tool calls vs JSON actions through TaskExecutor) for future optimization work.
WIP
Caution: Documentation below may be out of date.
🧠 Cognitive Architecture
AIRI's Minecraft agent is built on a four-layered cognitive architecture inspired by cognitive science, enabling reactive, conscious, and physically grounded behaviors.
Architecture Overview
graph TB
subgraph "Layer A: Perception"
Events[Raw Events]
EM[Event Manager]
Events --> EM
end
subgraph "Layer B: Reflex (Subconscious)"
RM[Reflex Manager]
FSM[State Machine]
RM --> FSM
end
subgraph "Layer C: Conscious (Reasoning)"
ORC[Orchestrator]
Planner[Planning Agent (LLM)]
Chat[Chat Agent (LLM)]
ORC --> Planner
ORC --> Chat
end
subgraph "Layer D: Action (Execution)"
TE[Task Executor]
AA[Action Agent]
Planner -->|Plan| TE
TE -->|Action Steps| AA
end
EM -->|High Priority| RM
EM -->|All Events| ORC
RM -.->|Inhibition Signal| ORC
ORC -->|Execution Request| TE
style EM fill:#e1f5ff
style RM fill:#fff4e1
style ORC fill:#ffe1f5
style TE fill:#dcedc8
Layer A: Perception
Location: src/cognitive/perception/
The perception layer acts as the sensory input hub, collecting raw signals from Mineflayer and turning them into higher-level, rate-limited perception events.
Pipeline:
- Mineflayer listeners collect raw perception events (sight/hearing/felt), including distance and line-of-sight when applicable.
- Raw events are queued in a buffer and drained on the cognitive tick.
- An attention detector aggregates events via leaky buckets and emits attention/perception events only on threshold crossing (e.g. sustained movement, punching, teabagging, interesting sounds).
Key files:
mineflayer-perception-collector.tsraw-events.tsraw-event-buffer.tsattention-detector.tspipeline.ts
Layer B: Reflex
Location: src/cognitive/reflex/
The reflex layer handles immediate, instinctive reactions. It operates on a finite state machine (FSM) pattern for predictable, fast responses.
Components:
- Reflex Manager (
reflex-manager.ts): Coordinates reflex behaviors - Inhibition: Reflexes can inhibit Conscious layer processing to prevent redundant responses.
Layer C: Conscious
Location: src/cognitive/conscious/
The conscious layer handles complex reasoning, planning, and high-level decision-making. No physical execution happens here anymore.
Components:
- Orchestrator: Coordinates "Thinking" vs "Chatting" tasks.
- Task Manager: Manages concurrent Primary (Physical) and Secondary (Mental) tasks.
- Planning Agent: pure LLM reasoning to generate plans.
- Chat Agent: Generates natural language responses.
Layer D: Action
Location: src/cognitive/action/
The action layer is responsible for the actual execution of tasks in the world. It isolates "Doing" from "Thinking".
Components:
- Task Executor: Receives a
Planand executes it step-by-step. Handles retry logic and errors. - Action Agent: The interface to low-level Mineflayer skills (move, place, break).
🔄 Event Flow Example
Scenario: "Build a house"
Player: "build a house"
↓
[Perception] Event detected
↓
[Conscious] Architect plans the structure
↓
[Action] Executor takes the plan and manages the construction loop:
- Step 1: Collect wood (calls ActionAgent)
- Step 2: Craft planks
- Step 3: Build walls
↓
[Conscious] ChatAgent confirms completion: "House is ready!"
📁 Project Structure
src/
├── cognitive/ # 🧠 Perception → Reflex → Conscious → Action
│ ├── perception/ # Event ingestion
│ │ ├── mineflayer-perception-collector.ts
│ │ ├── raw-events.ts
│ │ ├── raw-event-buffer.ts
│ │ ├── attention-detector.ts
│ │ └── pipeline.ts
│ ├── reflex/ # Fast, rule-based reactions
│ │ └── reflex-manager.ts
│ ├── conscious/ # LLM-powered reasoning
│ │ ├── blackboard.ts # Shared working memory
│ │ ├── brain.ts # Core reasoning loop/orchestration
│ │ ├── completion.ts # LLM completion helper
│ │ ├── handler.ts # Routes stimuli into the brain
│ │ ├── task-manager.ts # Manages concurrent tasks
│ │ ├── task-state.ts # Task lifecycle enums/helpers
│ │ └── prompts/ # Prompt definitions (e.g., brain-prompt.ts)
│ ├── action/ # Task execution layer
│ │ ├── task-executor.ts # Executes planned steps with retries
│ │ └── types.ts
│ ├── container.ts # Dependency injection wiring
│ ├── index.ts # Cognitive system entrypoint
│ └── types.ts # Shared cognitive types
├── agents/ # Specialized agents
│ ├── action/ # Low-level actuator bridge
│ ├── planning/ # Goal planner (LLM)
│ ├── chat/ # Conversational responses
│ └── memory/ # Memory-related helpers
├── libs/
│ └── mineflayer/ # Mineflayer bot wrapper/adapters
├── skills/ # Atomic bot capabilities
├── composables/ # Reusable functions (config, etc.)
├── plugins/ # Mineflayer/bot plugins
├── web/ # Debug web dashboard
├── utils/ # Helpers
├── debug-server.ts # Local debug server entry
└── main.ts # Bot entrypoint
🎯 Design Principles
- Separation of Concerns: Each layer has a distinct responsibility
- Event-Driven: Loose coupling via centralized event system
- Inhibition Control: Reflexes prevent unnecessary LLM calls
- Extensibility: Easy to add new reflexes or conscious behaviors
- Cognitive Realism: Mimics human-like perception → reaction → deliberation
🚧 Future Enhancements
-
Perception Layer:
- ⏱️ Temporal context window (remember recent events)
- 🎯 Salience detection (filter noise, prioritize important events)
-
Reflex Layer:
- 🏃 Dodge hostile mobs
- 🛡️ Emergency combat responses
-
Conscious Layer:
- 💭 Emotional state management
- 🧠 Long-term memory integration
- 🎭 Personality-driven responses
🛠️ Development
Commands
pnpm dev- Start the bot in development modepnpm lint- Run ESLintpnpm typecheck- Run TypeScript type checkingpnpm test- Run tests
🙏 Acknowledgements
🤝 Contributing
Contributions are welcome! Please feel free to submit a Pull Request.