## Architecture

**Request pipeline** (`src/cya/orchestrator.py`):
1. Collect local context (`context/collector.py`)
2. Recall memory via phase-memory ports (`memory/__init__.py`)
3. Classify risk (`safety/risk.py`) — rule-based; memory signals add caution only
4. Call LLM via `LLMAdapter` Protocol (`llm/adapter.py`) — FakeLLMAdapter today
5. Render explainable response (Rich)

**Memory (Profile 0 + Profile 1 spike):**
- User-controlled local JSON behind explicit ports (`remember`, `recall`, `forget`, `export`)
- Kinds: `preference`, `retrospection`, `interaction_goal`, `reflection`
- Directory/project-bound activation via `activation_context`
- `cya retrospect` feeds higher-order memory; optional verbal lesson capture (Profile 1)

**Boundaries:** See `repo-boundary.md`. No production path bypasses the adapter or memory ports.

## Quick Reference

- Memory contract: `MemoryVision.md`
- Activation/retrospection concept: `docs/cya-memory-activation-and-retrospection-concept.md`
- phase-memory feedback: `docs/phase-memory-optimization-suggestions.md`
- `~/state-hub/mcp_server/TOOLS.md` — MCP tool reference