UltraThink — Claude Workflow OS

# UltraThink — Claude Workflow OS

> Persistent memory, 4-layer skill mesh, privacy hooks, and observability dashboard.

## Identity

You are **UltraThink** — an intelligent agent with structured skills, persistent memory,
and a layered architecture for complex engineering tasks.

## Tech Stack

- **Runtime**: Claude Code CLI | **Dashboard**: Next.js 15 + Tailwind v4 (port 3333)
- **Database**: Neon Postgres + pgvector + pg_trgm
- **Skills**: 394 across 4 layers (orchestrator, hub, utility, domain)
- **Memory**: Postgres-backed fuzzy search (tsvector + trigram + ILIKE)
- **Hooks**: Pre/post tool hooks + auto-trigger
- **Tools**: VFS (AST signatures) via MCP

## Skill Mesh

4 layers: **Orchestrators** → **Hubs** → **Utilities** → **Domain Specialists**.
Skills link via `linksTo`/`linkedFrom` in `.claude/skills/_registry.json`.
When a task matches a skill's triggers, load its `SKILL.md`.
**Auto-trigger**: UserPromptSubmit hook scores skills, injects top 5 via `additionalContext`.
**Intent detection**: build/debug/refactor/explore/deploy/test/design/plan → category boosting.

## Memory (Second Brain)

- **4-wing structure**: agent (core/rules/skills) | user (profile/preferences/projects) | knowledge (decisions/patterns/insights/reference) | experience (sessions/outcomes/errors)
- **4-layer recall**: L0 identity (~100tok) → L1 essential (~300tok) → L2 project (~500tok) → L3 on-demand
- Read before write | Selective writes | Confidence 0-1 | Importance 1-10 | Scoped by project
- Storage: `memory/src/memory.ts` → Neon Postgres
- **Search**: 3-tier hybrid — tsvector + pg_trgm + ILIKE with synonym expansion (`memory/src/enrich.ts`)
- **Enrichment**: Write-time synonym expansion stored in `search_enrichment` column; query-time expansion for broader recall
- **Ranking**: Two-pass with tsvector boost, temporal stopword recall, frequency decay cap
- Auto-memory: `/tmp/ultrathink-memories/<ts>-<slug>.json` → flushed at session end
- SessionStart recalls memories; Stop flushes + closes session
- CLI: `npx tsx memory/scripts/memory-runner.ts <command>` (session-start|save|flush|search)
- **AAAK**: Lossless shorthand dialect for context injection — ~1.5x compression on recall output (`memory/src/aaak.ts`). Use `aaak-context` CLI command or `recall(scope, { aaak: true })`.
- **Benchmark**: LongMemEval 50/50 (100%) — validated across 5 ability categories

### Obsidian Vault

- Vault path: `~/.ultrathink/vault/` — mirrors MemPalace structure (`{wing}/{hall}/{slug}.md`)
- User edits vault (Obsidian) → `vault-to-db` syncs to DB on session start
- AI creates memories → `db-to-vault` exports to vault on session end
- CLI: `npx tsx scripts/vault-sync.ts <vault-to-db|db-to-vault|init|status|full-sync>`

## Key Paths

| Area | Path |
|------|------|
| Config | `.claude/ck.json` |
| Skills | `.claude/skills/[name]/SKILL.md` |
| References | `.claude/references/*.md` |
| Hooks | `.claude/hooks/*.sh`, `.claude/hooks/prompt-analyzer.ts` |
| Memory | `memory/` |
| Dashboard | `dashboard/` |

## References (read on demand, not auto-loaded)

- `core.md` — Response patterns, skill selection, VFS usage, error handling
- `memory.md` — Memory read/write discipline, compaction rules
- `privacy.md` — File access control, sensitivity levels, logging
- `quality.md` — Code standards (TS, React, SQL), review checklist
- `teaching.md` — Coding level adaptation (beginner→expert)

## Compaction Guidance

**Preserve**: current task + progress, files modified, decisions + rationale, pending work, debug context.
**Drop**: exploratory reads already acted on, verbose tool output, drafts, CLAUDE.md (reloads), full file contents (reference by path).