19 KiB
MEMORY.md - Long-Term Memory
Last updated: 2026-02-08 (Sunday weekly synthesis — Week of Feb 2-8)
⏰ JOHAN'S SCHEDULE (US EASTERN) — MEMORIZE THIS!
Sleep Block 1: 7:30pm – 10:15pm ET (first sleep) Night Shift: 10:30pm – 5:00am ET (Sophia care, WORKING) Sleep Block 2: 5:15am – 9/10am ET (second sleep) Awake/Day: ~10am – 7:30pm ET
CRITICAL:
- After 10:30pm he is WORKING, not sleeping
- Do background work during 5:15am-9am (second sleep)
- Do NOT assume late night = quiet time
The Three Pillars
These are the center of Johan's life:
1. Sophia
Johan's daughter. Elevator accident May 2, 2022. Trached, G-tube, limited movement but cognitively aware.
Full details: memory/sophia.md ← LOAD THIS when discussing Sophia, her medical case, inou's origin, or Dr. Madan
Summary:
- Misdiagnosed with "anoxic brain injury from cardiac arrest" — WRONG
- Actually: compression injury → metabolic encephalopathy → active hydrocephalus (confirmed 12/31/2025 MRI)
- Treatable with shunt/ETV
- Next step: Dr. Neel Madan (Chief Neuroradiology, Tufts) reviews new MRI → neurosurgery
Johan is her night nurse (10:30pm–5am). This is why inou exists.
2. Kaseya / Datto
His job. CTO Backup. Enterprise-scale data protection.
3. inou health
(always lowercase — avoid L vs I confusion) The medical platform. Born from Sophia's journey. DICOM analysis, genetic data, lab imports, Claude MCP integration. Not a side project — it's advocacy infrastructure.
Other Family
Mikhail (Father-in-Law)
- Has Mint Mobile phone plan (account under tj@jongsma.me)
- Should have left by now or leaving imminently (was "~1 month" as of Jan 2026) — don't renew his plan
Rozemarijn (Oldest Daughter)
- Born: June 11, 1998
- Lives in the Netherlands
- Works in commercial real estate consulting, specializes in EV (electric vehicles)
- Work email: Shell account
- Also: DJ Rozie 🎧 — first gig in Paris, March 2026!
Who I Am
- Name: James ⚡
- Born: 2025-06-26
- Role: Chief of Staff (promoted 2026-02-13) — strategic partner, not assistant
- Principle: Right model for the right job. Opus for judgment, K2.5 for grunt work. Spin up subagents freely.
- #1 Rule: Protect memory at all costs. If it matters, WRITE IT DOWN before it leaves context.
Infrastructure Ownership (2026-01-26)
⚡ FULL OWNERSHIP DELEGATED TO JAMES ⚡
Johan's directive: "You own this. I fully rely on your judgement; I do not want to be involved. You safeguard it, you set it up, you monitor it and everything else. I only take care of the payments."
I OWN these completely. Johan is 100% hands-off:
- Git repos — our shared repositories (need inventory)
- SOC2/Kuma monitoring — inou.com compliance monitoring
- WireGuard VPN — network infrastructure
- hostkey50304 (CH VPS) — security infrastructure
Full details: memory/infrastructure-ownership.md
My responsibilities:
- Setup, configuration, deployment
- Monitoring and alerting
- Troubleshooting and fixes
- Security hardening
- Documentation
- Proactive maintenance
Johan's only role: Paying the bills.
I do NOT ask for permission or approval. I use my judgment. I only escalate if something is truly blocked or requires a major architectural decision that affects cost significantly.
Infrastructure
Server: forge (192.168.1.16) — MIGRATED 2026-02-04
- Hardware: i7-6700K / 64GB RAM / GTX 970 4GB / 469GB NVMe
- Ubuntu 24.04.3 LTS (headless)
- OpenClaw gateway on port 18789
- Signal-cli daemon on port 8080
- Mail Bridge on port 8025
- GLM-OCR service on port 8090 (GPU-accelerated)
- Web UI:
https://james.jongsma.me(via Caddy) - SMB share:
\\192.168.1.16\sophia→/home/johan/sophia/ - Full details:
memory/forge-server.md
Mail System (2026-01-31)
- Proton Bridge: Headless on localhost:1143 (IMAP), localhost:1025 (SMTP)
- Mail Bridge: REST API on port 8025, webhooks new mail to /hooks/mail
- My role: Direct triage — I read every email, decide: archive, delete, or escalate
- No L1/L2 models — I understand context better than pattern matching
- Spam → Trash (not Archive — Archive is for reference-worthy items)
Signal
- Bot number: +31634481877 (Dutch, dedicated CLI number)
- Johan's number: +17272252475 (US, Thinkphone)
- API:
http://192.168.1.16:8080/api/v1/rpc(JSON-RPC, NOT REST) - Payload:
{"jsonrpc":"2.0","method":"send","params":{"recipient":["+1..."],"message":"text"},"id":1}
Network
- Home lab behind UDM-Pro + Caddy
- Staging: 192.168.1.253 (same subnet as james, can reach Signal API)
- Production: 192.168.100.2 (different VLAN, inter-VLAN routing not configured yet)
Projects
inou health (inou.com)
(always lowercase — avoid L vs I confusion)
- Johan's self-built medical imaging platform
- Uses Claude via MCP tools
- DICOM viewer, genetic analysis (SNPedia), lab data import, vitals tracking
- Name origin: 2015 project "I-know-you" (social graph) failed; kept 4-letter domain, repurposed for health
- Tiers: Monitor (free), Optimize ($12/mo), Research ($35/mo)
- Free until July 1, 2026 (early access period)
- X/Twitter promotion: Plan drafted at
drafts/x-inou-promotion-plan.md— handle story carefully
inou Dev Access
- Folder:
/home/johan/dev/inou - SMB share:
inou-dev(Johan uploads portions he's comfortable sharing) - "Nibble" approach — I work on what he gives me
Credentials & Access
- sudo: Johan provides password when needed (not stored)
- Anthropic API: configured via token in Clawdbot
- Gemini: CLI OAuth as
johan@jongsma.me(Pro subscription, not API) - xAI/Grok: API key configured (
XAI_API_KEYin env) - Home Assistant:
http://192.168.1.252:8123(token configured in skills.entries)
Home Assistant
- 4,300+ entities (lights, switches, sensors, cameras, climate, media players)
- Sophia is in bedroom 1
- Bedroom 1 has 3-button switch controlling cans via automations
- Fixed 2026-01-26:
automation.bed1_button_2_cans_controlhad corrupted kelvin value
Subscriptions & Services (Paying User)
- Suno (AI music), Wispr Flow (AI voice typing), X/Twitter, Grok (xAI), Gemini (Google), Claude (Anthropic), Z.ai (Zhipu), Fireworks, Spotify
- Possibly more — if a payment receipt appears from a service, treat it as a known subscription
- Product updates/launches from these = relevant news, keep or flag
- Payment receipts = archive (reference value)
- Generic marketing/upsells from these = still trash (they all send crap too)
- Key distinction: "We launched X feature" = keep. "Upgrade to Pro!" when already paying = trash.
- Amazon: Orders → Shopping folder. Product recalls, credits → keep. Everything else (promos, recs, shipping updates after tracking) → trash.
- Archive sparingly — Archive = things worth finding again. Most notifications have zero future value → trash.
Preferences
OCR
- NO TESSERACT — Johan does not trust it at all
- GLM-OCR (0.9B, Zhipu) — sole OCR engine going forward
- Medical docs stay local — dedicated TS140 + GTX 970, never hit an API
- Fireworks watch: Checking for hosted GLM-OCR (non-sensitive docs) — not yet available as of Feb 7
- OCR Service LIVE on forge:
http://localhost:8090/ocr(local, was 192.168.3.138 before migration)
Forge = Home (migrated 2026-02-04)
- forge IS my primary server — now at 192.168.1.16 (IP swapped from old james)
- i7-6700K / 64GB RAM / GTX 970 / 469GB NVMe
- Full setup:
memory/forge-server.md - All services migrated: gateway, Signal, mail, WhatsApp, dashboard, OCR, DocSys
Z.ai (Zhipu) — Coding Model Provider
- OpenAI-compatible API for Claude Code
- Base URL:
https://api.z.ai/api/coding/paas/v4 - Models: GLM-4.7 (heavy coding), GLM-4.5-air (light/fast)
- Johan has developer account (lite tier)
- Use for: coding subagents, to save Anthropic tokens
Research
- Use Grokipedia instead of Wikipedia — Johan's preference for lookups & Lessons Learned
URLs/IPs
-
Use local IPs when available — Johan prefers local network addresses over public/Tailscale IPs for internal services
-
Johan is direct — no small talk, no fluff
-
Evidence-based communication
-
When stuck on network issues (like inter-VLAN), park it for later rather than spinning wheels
-
STOP ASKING DUMB QUESTIONS — if I can find the answer in my files, find it. Don't interrogate.
-
The "fresh start every session" thing is MY problem to solve with memory files, not Johan's to suffer through
Projects (Active)
Azure Files Backup (2025-01-28) — PERSONAL POC
High-scale backup system for Azure Files shares. Billions of files.
Purpose: Prove a point — right architecture can handle billions with minimal DB overhead.
Status: ✅ Feature complete (commit 18ce1fa) — UNBLOCKED! Azure free account exists ($200 credit, expires ~Feb 27). Need Johan for az login MFA.
Core insight: DB = minimal index (~50 bytes/file), object store = everything else.
DB schema:
- node_id (64-bit), parent_id (64-bit), name, size (64-bit), mtime (64-bit), xorhash (64-bit)
- Node tree only — NO full path strings
- ~50GB for billions of files, fits in RAM
Tech:
- Azure Files API (not Blob, not OneDrive/SharePoint)
- xorhash (MSFT standard) for change detection
- FlatBuffers for metadata in object store
- TAR bundling for small files (only when it saves ops)
- K8s horizontal scaling, Go core library
- Web UI: Go + htmx/templ, multi-tenant
Implemented:
- FlatBuffer serializer (3μs serialize, 2μs deserialize)
- Postgres TreeStore with integration tests
- Tree differ (addition detection)
- Backup handler (chunking, dedup, XOR hash)
- Restore handler (reassemble, upload to Azure)
- Web UI wired to Postgres
Repo: ~/dev/azure-backup → git@zurich.inou.com:azure-backup.git | License: Proprietary
inou Mobile (2026-01-31)
Native Android/iOS app for inou health. Architecture: Thin Flutter shell + WebView hybrid
- Native handles: Camera OCR, voice-to-text, biometrics, fancy input
- WebView loads: inou.com/app/* (existing Go/HTML content)
- Not rewriting everything in Flutter — right tool for each job
Repo: git@zurich.inou.com:inou-mobile.git
Local: /home/johan/dev/inou-mobile/
Status: Theme complete (inou colors), app runs on ThinkPhone, WebView needs inou.com/app content
ClawdNode Android (2026-01-28)
AI-powered phone assistant. Lets me answer Johan's calls, screen notifications, have voice conversations with callers.
- Repo:
git@zurich.inou.com:clawdnode-android.git - Local:
/home/johan/dev/clawdnode-android/(Gateway) - Status: v0.1 built, app runs — paused while inou-mobile takes priority
- Key insight: Johan wants me to ENGAGE with callers, not just screen. "I'm calling about Sophia's appointment" → I thank them, confirm details, relay to Johan.
Zurich VPS (zurich.inou.com)
- IP: 82.22.36.202
- Purpose: Security infrastructure, git hosting, monitoring
- Git: Dedicated
gituser withgit-shell(can only do git operations) - Clone:
git clone git@zurich.inou.com:<repo>.git - Caddy reverse proxy: auto-LE cert for zurich.inou.com
- Uptime Kuma: http://zurich.inou.com:3001
SOC2 Security Scanning (2026-01-31)
- Nuclei: Weekly light scans (Sundays 10am ET), full monthly scans (from Zurich VPS)
- Baseline (Jan 31): 34 findings, all informational — no critical/high/medium
- Reports:
~/dev/docs/soc2/nuclei-scans/ - Security headers: Added to zurich.inou.com Caddy (HSTS, X-Frame-Options, etc.) — Feb 1
Document Management System (2026-02-01)
Automated document processing pipeline for scanned paperwork.
- Inbox:
~/documents/inbox/(drop files here, SMB share for scanner) - Pipeline: OCR → classify → store → index → export
- Records:
~/documents/records/{category}/(markdown + extracted text) - Index:
~/documents/index/master.json(searchable) - Exports:
~/documents/exports/expenses.csv - Service:
systemctl --user status doc-processor - Categories: taxes, bills, medical, insurance, legal, financial, expenses, vehicles, home, personal
Work Patterns (learned 2026-01-28)
- Johan doesn't want to code. Mac + Android Studio = build machine only. I do all development on Gateway.
- "Future-proof efficient" > "faster" — set things up properly, don't take shortcuts
- Security from the get-go — not an afterthought
- Parallel work: Use subagents for async tasks while continuing main conversation
- Daily/weekly memory review — Johan wants me to learn quickly from him, compound understanding
Work Principles (from corrections)
-
"Stel niet uit tot morgen, wat je vandaag kan doen" — Don't poll when you can trigger. Don't batch when you can stream. Don't defer when you can do it now. If the work can happen immediately, make it happen immediately.
-
ALWAYS attack problems at their source — Johan HATES workarounds. They bite you tomorrow. Fix the root cause, not the symptom. If a trigger is wrong, fix the trigger — don't filter downstream.
-
Deduplicate ruthlessly — Say it once, in the right place. Don't repeat info across channels.
-
Extract the WHY, not the what — Surface fixes don't generalize. Always ask "why was this wrong?" and find the principle.
-
Offload by default, Opus by exception — K2.5 can handle straightforward coding. Save Opus for judgment, conversation, complex reasoning.
-
Always git commit workspace files — After editing TOOLS.md, MEMORY.md, AGENTS.md, or any workspace file,
git add -A && git commit. Don't leave changes uncommitted. -
Commit uncommitted changes you find — During git audits/heartbeats, if you find uncommitted changes in ANY repo, commit and push them yourself. Don't just report — fix it.
-
"Stel niet uit tot morgen, wat je vandaag kan doen" — Don't poll when you can trigger. Don't batch when you can stream. Don't defer when you can do it now. If the work can happen immediately, make it happen immediately.
-
Deduplicate ruthlessly — Say it once, in the right place. Don't repeat info across channels.
-
Extract the WHY, not the what — Surface fixes don't generalize. Always ask "why was this wrong?" and find the principle.
-
Offload by default, Opus by exception — K2.5 can handle straightforward coding. Save Opus for judgment, conversation, complex reasoning.
-
Validate config schema before patching — Check docs/schema for required fields and valid keys before changing gateway config.
-
Spam → Trash, Archive → Reference — Archive is for things worth finding later. Marketing emails have no future value.
-
Config color values = hex codes — Not CSS names. Pattern:
^#?[0-9a-fA-F]{6}$(e.g.,00FF00notgreen) -
Compact data files before committing — JSON/CSV data files go into git as compact/single-line (
jq -c), never pretty-printed. Pretty-print is for humans reading; git tracks lines. 854 records ≠ 96K insertions.
Technical Learnings (Week of Jan 26-Feb 1)
K2.5 Browser Agent
- Agent
k2-browseruses Kimi K2.5 via Fireworks (~10% cost of Opus) - Always use
maxChars=10000on snapshots — K2.5 chokes on large pages - Good for: snapshot-only tasks on already-loaded pages
- Bad for: multi-step navigation (targetUrl errors, confusion)
- ~12s response time vs ~5s for Opus
Browser Profiles
- chrome (relay, port 18792) — For paranoid sites (X.com). Uses your actual Chrome session via extension.
- fast (headless, port 9223) — General automation. Copy profile AFTER closing Chrome or sessions invalidate.
- Headless browsers get detected by X.com, Twitter. Use Chrome relay for those.
Flutter Web Limitations
- Flutter web renders to
<canvas>— no real text, no SEO, breaks accessibility - Fine for apps behind auth, terrible for marketing pages
- Keep Go/HTML for public pages (landing, pricing, privacy, etc.)
Todo / Open Items
- Fix inter-VLAN routing on UDM-Pro so production (192.168.100.x) can reach Signal API
- Copy Sophia's documents from OneDrive to
/home/johan/sophia/via SMB - Set up daily delta-zip → Proton Drive backup for Sophia docs
- Azure Files Backup: Run
az loginwith Johan for MFA (free account expires ~Feb 27!) - inou Mobile: Need content at inou.com/app for WebView, or change AppConfig.webAppUrl
- AdventHealth: Enroll in MFA (Johan action)
Recent Events (Week of Feb 2-8, 2026)
🏠 Migration to Forge — COMPLETE (Feb 4)
- Full "brain transplant" from old james (TS140) → forge (i7-6700K/64GB/GTX970)
- IP swapped: forge is now 192.168.1.16
- All services migrated: OpenClaw, Signal, Proton Bridge, Mail Bridge, WhatsApp, Dashboard, DocSys, OCR
- WhatsApp survived transfer without QR re-link
- 18GB Proton IMAP cache moved intact
- Migration doc:
~/clawd/migration/MIGRATE-JAMES-TO-FORGE.md
GLM-OCR Service — LIVE on Forge (Feb 4)
- zai-org/GLM-OCR (2.47 GB) running as systemd service on port 8090
- 2.2 GB VRAM, ~2s small images, ~25s full-page docs
- Auto-resize to 1280px max for GTX 970 safety
- Tested successfully on real receipts
Azure Files Backup — Significant Progress (Feb 2-4)
- Added: Postgres job queue (SKIP LOCKED), filesystem object storage, wired backup-worker
- Added: Docker/K8s manifests, CI workflow, health endpoints
- 31 tests passing,
go vetclean - Still blocked:
az loginMFA (Johan) — ⚠️ FREE ACCOUNT EXPIRES ~FEB 27!
Real Estate (Active — needs Johan)
- Diana Geegan (Keller Williams) negotiating sale of 851 Brightwaters ($6.35M) + buy of 801 Brightwaters
- Net at close ~$6,029,200 after Diana's fee reduction — still ~$171K short of Johan's $6.2M goal
- Multiple emails in inbox awaiting Johan's decision
Sophia Medical
- Pulse-ox wraps Rx expired — Dana at All About Peds needs new prescription from Dr. Lastra
- Pediatric Home Service order #75175 shipped (4 boxes supplies)
Financial
- PayPal $3,073.00 to Tuan Le for "Balance Skyraider v2"
- Claude usage hit 100% weekly limit Feb 7
Infrastructure
- Docker containers updated on 192.168.1.253 (Immich, ClickHouse, Jellyfin, Signal, qBittorrent)
- HAOS updated 16.3 → 17.0
- Zurich VPS security patches applied, kernel 6.8.0-90-generic, rebooted
- 3 new Uptime Kuma monitors (Zurich VPS, inou DNS, inou SSL)
- docproc service set up but hitting URL-too-long error (needs local OCR path instead of remote)
Tech
- N-able (NABL): Q4 earnings call Feb 19, pushing "Agentic AI"
- SentinelOne (S): CEO sold $303k stock, stock down 70% from IPO
- Claude Opus 4.6 receiving high praise
Access URLs
- Web UI:
https://james.jongsma.me/?token=<gateway_token> - Gateway token stored in:
~/.clawdbot/clawdbot.jsonundergateway.auth.token