Knowledge Architecture

SecondBrain + GBrain + PKA Vault + System Wiki — unified view as of 2026-04-21

What changed 2026-04-21 (Memory Unification Sessions 1–3): Auto-memory split from per-device (mac/luci/mike/) to per-persona (lucienne/luci/) · Auto-memories now enter GBrain (old invariant retired) · vault.db demoted to secondary, retiring mid-May (4-week parallel run) · Query order consolidated into canonical Rule 14 (4-step stack) · Extractor made portable: PKA_ROOT + PKA_MEMORY_PERSONA env vars · pka-sync.sh now imports curated + auto memories into GBrain · CLAUDE.md trimmed 68% (38KB → 12KB), verbose sections moved to wiki · New wiki pages: projects.md, cowork-integration.md

SecondBrain

~/cowork/SecondBrain/

Elmar's Obsidian vault — the raw knowledge lake. ~4,650 .md files. Capture first, synthesise later.

Key folders

Meetings/ 211+ dated notes (transcribed)
sources/ email-attachments, exco, family-docs, imports
wiki/ 73 compiled pages (projects + people)
ideas/, inbox/ quick capture

Auto-ingestion pipelines

exco-sync — Mac LaunchAgent, daily 10:00 SAST
sources-sync — Mac LaunchAgent, daily 10:15 SAST
email-attachment-sync — Luci, daily 7am SAST
dream cycle — Luci, 2am SAST (detect stale + recompile)
pka-sync — Mac, every 15 min (feeds GBrain)

Sync

Mac auto-commits every 15 min via LaunchAgent
Luci pulls every 15 min via cowork-sync task
Repo: conrelma/local-cowork (private)

GBrain PRIMARY

~/.gbrain/brain.pglite · 5,776 pages

PGLite + pgvector. Adopted 2026-04-17. Beats vault.db FTS on 4/5 classes, +18% nDCG@5, ~1000× faster. Now indexes everything including auto-memories.

A/B Scorecard

Person queriesGBrain +51% R@5

FinancialGBrain +46% R@5

TemporalGBrain +18%

Entity relationGBrain +40% R@5

Decision recallPKA +44% nDCG (synthesis)

Latency p95<50ms vs 38s

What GBrain indexes (5 sources) CHANGED

All SecondBrain .md files (~4,650 files, recursive)
PKA system wiki pages (type: system-doc)
PKA docs (type: project-doc)
Curated memories (Vault/memory/*.md) NEW
Auto-memories (auto/lucienne/ + auto/luci/) NEW

Integration

MCP: mcp__gbrain__query (primary tool)
Sync: com.gbrain.pka-sync LaunchAgent every 15 min
Retrieval: RRF hybrid (vector + keyword + multi-query)
Embed: OpenAI API (~/gbrain/.env), ~$0.02/machine

PKA Vault RETIRING MID-MAY

~/PKA/Vault/

vault.db is now secondary. 4-week parallel run, then Session 4 deletes index.py, schema.sql, vault.db. Markdown files remain source of truth.

Contents

vault.db SQLite — files, edges, FTS5 (retiring)
memory/ ~44 curated reference files
memory/auto/lucienne/ 206 session memories CHANGED
memory/auto/luci/ Luci sessions (populates next session) CHANGED
tasks/ YAML frontmatter tasks
workspace/ live running workspaces

Two-DB split (unchanged)

vault.db — Lucienne owns (rebuildable, retiring)
mc.db — Luci owns (Hetzner); tickets, task_runs, heartbeats
Vault MCP merges both via HTTP API

Session 4 (mid-May): what gets deleted

index.py, session_indexer.py, schema.sql, vault.db
vault_mcp.py shrinks to MC proxy (2 tools)
dashboard.py rewrites to flat files + GBrain

Auto-Memory CHANGED

Vault/memory/auto/{persona}/

Per-persona split (not per-device). Synced via git. Now ingested into GBrain — old "never enters GBrain" invariant retired 2026-04-21.

Persona directories

lucienne/

All Lucienne machines (Mac, XPS, Mac Mini, Cloud). autoMemoryDirectory: ./Vault/memory/auto/lucienne/

luci/

Luci on Hetzner. Absolute path: /home/lucienne/workspace/PKA/Vault/memory/auto/luci/

Write sources

Claude Code built-in memory (mid-session saves)
session-memory-extractor.py — SessionStart hook, headless Haiku
Env vars: PKA_ROOT, PKA_MEMORY_PERSONA NEW

GBrain integration NEW

pka-sync.sh imports both auto/lucienne/ and auto/luci/
Embedded alongside curated memories and wiki pages
Searchable via mcp__gbrain__query

System Wiki + Atlas

~/PKA/wiki/ · 18 pages

Atlas (Systems Architect) owns the wiki. First voice before PKA system changes, last voice after. Wiki pages indexed in GBrain as type: system-doc.

Wiki pages

pka-system.md system overview
memory-system.md architecture, invariant flip
auto-memory.md persona-split design
secondbrain-how-it-works.md 3-layer architecture
dashboard.md PKA dashboard vs Luci
team.md dispatch model, council, roster
projects.md SafairBru + LegalMind NEW
cowork-integration.md CoWork + SB patterns NEW
infrastructure.md, mission-control.md, repos-and-sync.md, ...

Atlas dispatch model

Brief — first voice before any PKA system change
Sign-off — last voice after dev loop; wiki updated before return
Explain — "how does X work?" dispatched to Atlas
Monthly review: first session of each month, all 4 council models

Query Order (Rule 14) CHANGED

Canonical 4-step stack — all retrieval follows this

Consolidated from 3 contradictory blocks into one canonical rule. GBrain is step 2 (primary); vault.db is step 3 (secondary, retiring).

Claude's native auto-memory — session context, MEMORY.md
GBrain — mcp__gbrain__query. SB, wiki, docs, curated + auto memories. Primary.
Flat-file / vault.db — mcp__vault__memory_search. Structured queries, stale checks. Secondary (retiring).
Luci MC API — tickets, heartbeats, task_runs, scheduler status.

Routing exceptions

PKA system questions → dispatch Atlas (Explain mode)
Complex multi-type recall → memory-manager skill

Skills (API Layer)

Interface between agents and vaults. Agents call skills, not edit markdown directly.

/brainQuick capture → classifies into SB

/meeting-notesAudio → transcript → structured note with wikilinks

/wiki-compilervault.db + sources → synthesized wiki pages (feeds GBrain)

/wiki-ingestPDF/DOCX → markdown in sources/ → indexed

/wiki-queryThin GBrain adapter — calls gbrain.query()

/memory-managerRetrieve-before-answer. Searches vault + GBrain before any claim.

/sources-syncYAML-driven folder watcher → SB as markdown

/email-indexSQLite on Luci, 2,026 emails, 23 projects

Actors & Transport

Who reads, who writes, how data moves.

Elmar

Browses SecondBrain via Obsidian. Messages Luci via Telegram when Lucienne is offline.

Lucienne (Mac)

Chief of Staff. Queries GBrain (primary). Owns vault.db (retiring). Dispatches Atlas, Luci, Scott, Tessa. Does not touch mc.db.

Luci (Hetzner)

Project Manager. Owns mc.db. Reads vault.db via git pull (read-only). Runs dream cycle, email-sync, scheduler. Her dashboard is separate.

Atlas (subagent)

Systems Architect. Owns ./wiki/. First + last voice on PKA system changes. Reads all wiki pages on every dispatch.

Git repos

conrelma/local-cowork → SecondBrain vault
conrelma/PKA → vault.db, memory/, wiki/ (15-min sync)
conrelma/claude → ~/.claude/ (vault, skills, hooks)

Key Data Flows

How information moves through the system (2026-04-21 state).

Capture

Elmar records meeting → Luci transcribes (Gemini) → /meeting-notes structures it → dream cycle recompiles wiki page overnight → pka-sync pushes to GBrain within 15 min.

SecondBrain/Meetings/YYYY-MM-DD-*.md

Retrieval

Elmar asks Lucienne about a project/person. She calls mcp__gbrain__query() → RRF hybrid returns ranked chunks with citations in <1s. No waterfall. No file-reading loop.

GBrain → compiled-truth pages score 0.90-0.97

Document Ingest

PDF/DOCX → markitdown or Gemini vision → markdown in sources/ → dream cycle compiles → pka-sync pushes to GBrain.

sources/ → SecondBrain → GBrain

Memory Creation CHANGED

Session ends → extractor runs at next SessionStart (Haiku) → writes to auto/{persona}/ → git sync propagates → pka-sync imports to GBrain. Searchable via gbrain.query() within 30 min.

auto/lucienne/ or auto/luci/ → GBrain

System Change

Atlas brief → dev loop (architect → developer → QA → reviewer) → Atlas sign-off → wiki updated → GBrain picks up wiki change on next sync.

./wiki/ updated before Atlas returns

Ops Data

Luci writes mc.db. Lucienne reads via Vault MCP activity_log + luci_status tools, which call MC API at port 3001 transparently.

mc.db → HTTP API :3001 → Vault MCP → Lucienne