refactor: replace 3-Sets diagnostic with focused attention filters and memory

The 3-Sets framework doesn't transfer well to agents — all three
dimensions are fully visible and controllable config, not hidden
human psychology. Removed the branding, kept the practical bits:

- attention-filters: what context each archetype receives (token savings)
- memory: persistent learnings across orchestrations (project knowledge)

skills/memory/SKILL.md

@@ -0,0 +1,60 @@
+---
+name: memory
+description: Use after orchestration to record learnings, and before orchestration to load project-specific context. Persistent knowledge that makes future orchestrations smarter.
+---
+
+# Persistent Memory
+
+ArcheFlow learns across orchestrations. After each run, extract what was discovered. Before the next run, load it so agents start smarter.
+
+## What to Record
+
+Store in `.archeflow/memory/project.md`:
+
+```markdown
+## Project Context
+- Language: TypeScript, strict mode
+- Package manager: pnpm
+- Test runner: vitest
+- CI: GitHub Actions (no DB access in CI)
+
+## Risk Map
+- auth/ — well-tested (95% coverage), normal risk
+- payment/ — no tests, elevated risk, Guardian should be thorough
+- legacy/api-v1/ — deprecated, do not modify
+
+## Model Notes
+- Type-heavy modules need Sonnet minimum (Haiku produced incomplete types)
+- Standard CRUD reviews work fine with Haiku
+
+## Shadow History
+- Explorer rabbit-holed in monorepo (cycle 1 of orchestration on 2026-03-28)
+  → added 10-file cap to Explorer prompt for this project
+```
+
+## What NOT to Record
+
+- Anything derivable from code or git history
+- Temporary state from current orchestration
+- Opinions or predictions
+- Anything that changes every week
+
+## When to Read
+
+Before orchestration, check if `.archeflow/memory/project.md` exists. If it does, inject relevant sections into agent prompts:
+
+- Explorer gets: Project Context
+- Guardian gets: Risk Map
+- Maker gets: Project Context
+- All agents get: Model Notes (for self-calibration)
+
+## When to Write
+
+After orchestration completes, update memory if anything new was learned:
+
+- New project conventions discovered by Explorer
+- Risk areas identified by Guardian
+- Model tier adjustments needed (Haiku insufficient, or Opus unnecessary)
+- Shadow patterns that recurred
+
+Update existing entries — don't append endlessly. Memory should be a current snapshot, not a changelog.