feat(arch): implement 'Code as Thought' architecture and formalize PSF Consensus Loop

projects/org-skill-memex/docs/ARCHITECTURE.org

@@ -0,0 +1,64 @@
+#+TITLE: Org-Agent Memex Architecture Notes
+#+AUTHOR: Amr
+#+CREATED: [2026-03-17 Tue]
+#+BEGIN_COMMENT
+Core architectural principles and design decisions for the org-agent memex system.
+#+END_COMMENT
+
+* Core Philosophy: Single User, Single Agent
+
+** Why This Scope?
+
+The system is deliberately designed for *one human, one AI assistant*:
+
+- **No coordination complexity**: One agent owns one workflow (Scribe = Atomic Notes (Zettelkasten) distillation, GTD Manager = task promotion)
+- **No conflict resolution**: Agent reads from immutable sources (daily logs) and writes to separate targets (atomic notes, GTD promotions)
+- **No multi-agent negotiation**: The assistant doesn't delegate to sub-agents; it executes skills directly
+
+This is *not* a multi-agent orchestration system. It's personal automation.
+
+* Generalization via Environment Variables
+
+** Principle: Build with generalization, keep variable values out**
+
+All identity-specific and configuration values live in `.env`:
+
+| Variable | Purpose |
+|----------|---------|
+| MEMEX_USER | The human user's name (e.g., "Amr") |
+| MEMEX_ASSISTANT | The AI assistant's identifier (e.g., "Agent") |
+| CURRENT_TEXT_MANIPULATION_MODEL | The LLM tier for text processing |
+| MEMEX_* paths | Folder structure (PARA hierarchy) |
+
+Skills reference these as `$VARIABLE` in scripts or get instructed to use them. No hardcoded names in skill logic.
+
+* Source of Simplicity
+
+** What makes this project tractable:**
+
+1. *Standing on established frameworks*: Org-mode, Atomic Notes (Zettelkasten) method, GTD, PARA organization—the hard thinking is already done
+2. *Git as state machine*: Rather than building custom sync or consensus, we use Git commits as the source of truth for "what's new"
+3. *Immutable sources*: Daily logs are append-only; the Scribe never writes to them
+4. *Deterministic outputs*: Atomic notes have clear rules (concept-filenames, id: backlinks, no dates in names)
+
+** What we're NOT building** (which would add complexity):
+- Multi-user collaborative editing
+- Real-time synchronization across devices
+- Agent-to-agent task delegation protocols
+- Distributed state management
+- Conflict resolution for simultaneous edits
+
+The complexity is in the *workflow logic*, not the technical infrastructure.
+
+* Future: Linking with Native Org-Agent
+
+** Phase 1** (current): OpenClaw orchestrates cloud LLMs using SKILL.md definitions
+** Phase 2** (future): Native `org-agent` (Common Lisp) executes the same skills locally
+
+The interface remains constant:
+- Skill definitions in Org-mode format (SKILL.md)
+- .env configuration
+- PARA folder structure
+- Git-based state tracking
+
+When `org-agent` matures, it can read and execute the same skill files we're writing today. The transition from cloud-based to local inference becomes seamless because the *specification* (Org files) is implementation-agnostic.