Architecture

cognitive-engine is a monorepo with 14 packages, each handling one cognitive module.

Package Map

Layer	Package	Role
Core	@cognitive-engine/core	Types, interfaces, events, config
Core	@cognitive-engine/math	Linear algebra, sampling, temporal math
Perception	@cognitive-engine/perception	Input analysis, entity extraction, belief candidates
Memory	@cognitive-engine/memory	Episodic and semantic memory
Reasoning	@cognitive-engine/reasoning	World model, working memory, intention generation
Mind	@cognitive-engine/mind	Reflection, relationships, open loops, emotional triggers
Emotional	@cognitive-engine/emotional	Emotional state modeling
Social	@cognitive-engine/social	Rapport, boundaries, communication preferences
Temporal	@cognitive-engine/temporal	Patterns, causal chains, predictions
Planning	@cognitive-engine/planning	Multi-step plan generation
Metacognition	@cognitive-engine/metacognition	Contradiction detection, strategy tracking, cognitive load
Orchestrator	@cognitive-engine/orchestrator	Full pipeline orchestration
Bandit	@cognitive-engine/bandit	Thompson Sampling for strategy selection
Storage	@cognitive-engine/store-memory	In-memory store implementation
Provider	@cognitive-engine/provider-openai	OpenAI LLM and embedding provider

Data Flow

The orchestrator runs a cognitive cycle on each input:

1. Perceive     - analyze input, extract entities and intent
2. Remember     - query episodic + semantic memory for context
3. Reason       - form beliefs, generate intentions
4. Plan         - create multi-step plan if needed
5. Metacognize  - check for contradictions, adjust strategy
6. Respond      - generate output with full cognitive context
7. Store        - save episode to memory for future recall

Provider Interface

All LLM calls go through the LlmProvider interface:

interface LlmProvider {
  complete(messages: LlmMessage[], options?: LlmOptions): Promise<LlmResponse>
  completeJson<T>(messages: LlmMessage[], options?: LlmOptions): Promise<LlmResponse & { parsed: T }>
}

Implement this interface to use any LLM. The @cognitive-engine/provider-openai package provides a ready-made OpenAI implementation.