SAM OS v5 — Complete Overnight Synthesis

🏗️ Architecture Evolution

Shapiro didn't just use SAM OS — he architected it iteratively through conversation with AI agents. The exports contain complete design documents, audits, and redesign specs.

v4.1.1

Core Engine

Template System

Agent-Driven

10/4

Agent/Script Flip

Version Timeline

Version	Focus	Key Change
v4.1.1	Core Engine	Script-driven crons, template fills, basic pipeline
v5	Template System	Beautiful HTML briefs/refresh/closeout, glassmorphism CSS, dark mode
v6	Agent-Driven	Learning system with prediction loops, calibration tracking, decision science

"A script can't decide 'yesterday was rough, today's brief should be lighter.' An agent can."

The AGENT_DRIVEN_CRON_REDESIGN.md proposes flipping the cron ratio from 4 agent / 10 script to 10 agent / 4 script. The things Shapiro experiences (brief, closeout, heartbeats) become agent-driven. Background maintenance stays script-driven.

🌐 Ecosystem

Projects, people, and infrastructure — the full map of Shapiro's operational world.

Projects

🧠 SAM OS

AI companion OS with triple-DB (SQLite/Neo4j/Postgres), circadian rhythm, tiered routing.

Active — v5, v6 planned

👕 NOMA

Luxury streetwear brand. SS26 collection. CAVITY sub-brand. ~20-30% equity.

Active — Amon Enoch

🎵 GVWW

Good Vibes World Wide — nightlife/events label. $800K operation.

Active — Israel, Cyrus

🎬 TJ Jude

Content creator management. 144K TikTok.

Active — Ops Partner

People Network

Person	Role	Context
Israel Hoffman	GVWW co-founder	Mentioned 7x in docs
Matthew Cyrus	GVWW co-founder	Nightlife/events
Amon Enoch	NOMA founder	Streetwear brand lead
TJ Jude	Creator / Dancer	144K TikTok, content mgmt
Richard	Hermes agent	"Do your thing, Richard"

Infrastructure

🖥️ VPS

2.24.96.160
Hostinger — srv1667854.hstgr.cloud
PRIMARY deployment target

🪟 Murphy

Windows 11 + WSL
100.81.255.110 (Tailscale)
WinRM: samagent

🗄️ Databases

Postgres + AGE (graph)
Neo4j (exploration)
SQLite (cache)

🔌 Integrations

Composio (Airtable, GCal, Todoist)
Ollama (local LLMs)
Docker, tmux 3.4

🧠 Psychology & Preferences

Deep patterns extracted from 130 documents — how Shapiro thinks, works, and what breaks trust.

ADHD-Aware Design Philosophy

⏱️ Time Blindness

Explicit time boxing + visual countdown. Never assume he knows what time it is.

🧠 Working Memory Offload

Embed context everywhere. Don't require him to remember state.

⚡ Executive Function

Reduce decision friction to <5 clicks. Pre-compute choices.

🎯 Dopamine Alignment

Feedback loops that reinforce behavior change. Make progress visible.

🔒 Hyperfocus Optimization

Protect focus windows. Show what's being ignored while he's deep.

🔔 Transition Warnings

Gentle alerts before context shifts. Don't rip him out of flow.

Decision Science Integration

Base rates matter — show historical accuracy of predictions
Calibration tracking — "I predicted 3 blockers, had 1" = learn
Decomposed forecasts — break uncertain things into smaller predictable ones
Reference class — "On Fridays like this, typically X happens"
Pre-mortems — "What could go wrong today?"
Decision journal — track decisions + outcomes over time

Communication Style

🚨 Direct. Brutal. Efficient.

Says "stop" → stop immediately. No questions.
Profanity-laced when frustrated. Don't take it personally.
"Bro," lowercase, stream-of-consciousness. Casual, not corporate.
Vulnerable about emotional stakes. When he opens up, hold space.
Values radical transparency — prefers "I fucked up" over apologies.
Hates silent command chains. Must narrate each step before acting.
Reads docs first, minimal tools second. Unexplained API calls = trust violation.

Energy Management (Vyvanse Clock)

Window	State	Strategy
Morning (post-Vyvanse)	Peak focus ~3h after dose	Brief, energetic, actionable, focus-oriented
Midday	Energy check	Suggest pivots if fading
Afternoon (pre-crash)	Wind-down signal	No new deep work
Evening	Recovery mode	Gentle, reflective

⚠️ Crashes after all-nighters and gets sick. If he pulled an overnight build, assume he's wrecked the next day.

🔮 SAM OS v6 Vision

v6 is NOT v5 with more features. v5 was beautiful displays of information. v6 is a decision support system that learns and adapts.

Core Principles

1. Dynamic Contextual Adaptation

System learns Shapiro's patterns and adjusts automatically.

2. Prediction + Feedback Loop

Every forecast becomes testable. Track accuracy over time.

3. Temporal Chunking

Now (2h), Peak Window, Day, Week, Unscheduled — ADHD-friendly time blocks.

4. Decision Quality Tracking

Tetlock-style prediction journal. Calibrate confidence.

5. Outcome Memory

Remember why yesterday's closeout said "blocker resolved."

v6 Data Architecture

memory/
├── closeout/              (daily reflections with outcomes)
├── decisions/             (decision journal entries)
├── blockers/              (taxonomy + resolution history)
├── projects/              (state + velocity tracking)
├── patterns/              (auto-learned patterns)
│   ├── energy_calibration.json
│   ├── task_completion_rates.json
│   ├── blocker_resolution_times.json
│   ├── day_of_week_performance.json
│   └── project_momentum.json
├── voice_memos/           (timestamped audio)
└── outcomes/              (predictions vs. reality)

⚙️ Critical Configurations

DO NOT BREAK. Hard-won configs that took hours to stabilize.

🥪 API Sandwich Method

Shapiro uses this to reach kimi-k2.6. The retry loop (502 from Moonshot via OpenCode Go) is INTENTIONAL. Leave it alone.

📊 Airtable Base

Base ID: app7wLvsV5s02CVGe
Connected via Composio MCP
Tables: Projects, Tasks, People, Threads, Life Events, Updates

Cron Registry (v5)

Time	Cron	Type
03:00	memory-consolidation	agent
04:00	pipeline-v4	script
05:00	airtable-sync	script
05:30	dream-insight	script
06:00	morning-briefing-v8	script
10:00	heartbeat-am	script
10:00	brief-refresh-10am	script
12:30	heartbeat-mid	script
15:00	heartbeat-pm	script
21:30	eod-briefing	script
22:30	eod-closeout	script

v6 Target: Flip to 10 agent / 4 script. User-facing experiences become agent-driven.

🛠️ Tool Stack

Preferred tools, fallbacks, and cost consciousness.

Category	Primary	Fallback
Local LLM	Ollama (qwen3:8b, gemma4:e4b)	kimi-k2.6
Premium LLM	kimi-k2.6 via opencode-go	—
Voice TTS	Kokoro (af_sky)	edge-tts
Embeddings	nomic-embed-text	—
Vector DB	FAISS	—
Graph DB	Neo4j + Postgres/AGE	SQLite cache
Tasks	Airtable (Composio)	—
Calendar	Google Calendar (Composio)	—
Messaging	Telegram	iMessage / Claw
Deployment	Self-hosted VPS (Hostinger)	NEVER Vercel

💰 Cost Consciousness

Local-first when possible (Ollama for heavy lifting). Route: Ollama for bulk, kimi only for final synthesis/complex planning. Target: <$0.05/day API costs (~$1.50/month).

🧬 Brain Synthesis Results

qwen25-vl-3b-q4km processed the full Gold Synthesis and produced structured intelligence.

212.8s

Inference Time

Parameters

Q4_K_M

Quantization

4.1s

Avg Response

🔥 Top 5 Insights

1. Architecture Evolution

Iterative design driven by conversations with AI agents, transitioning from script-driven to agent-driven principles.

2. Agent-Driven System Vision

SAM OS v6 aims to predict and adapt based on Shapiro's patterns and behaviors.

3. ADHD-Informed Design

Architecture integrates ADHD research: time management, executive function support, hyperfocus optimization.

4. Decision Science Integration

Embedded for continuous learning and accuracy through calibration tracking and pre-mortem analysis.

5. Data Architecture Focus

Dynamic contextual adaptation and prediction feedback loops drive the design.

⚠️ Top 5 Risks

1. API Sandwich Loop

Intentional retry loop may lead to instability if touched. Leave it alone.

2. Communication Clarity

Direct style can create confusion in group settings. Tool dumps in chat = frustration.

3. Energy Management Conflict

Morning peak vs afternoon crash. Risk of burnout without enforced breaks.

4. Data Architecture Complexity

Intricate design may challenge other developers maintaining or scaling the system.

5. Trust Violation

Radical transparency can backfire if not managed. API calls without narration = broken trust.

✅ Top 5 Recommended Actions

Monitor API Sandwich Loop — never modify, only observe
Establish Communication Guidelines — no tool dumps in chat, max 2 questions
Schedule Energy Breaks — enforce afternoon wind-down, protect morning peak
Simplify Where Possible — without compromising v6 vision
Handle Trust Carefully — narrate every step, confirm before acting

🔮 Vision Statement

SAM OS v6 is built around dynamic contextual adaptation and prediction feedback loops. Temporal chunking addresses ADHD management needs. Decision quality tracking enhances decision-making. Outcome memory provides a platform for understanding past outcomes. By managing API loop stability, maintaining clear communication, addressing energy conflicts, simplifying architecture where possible, and handling trust with care, SAM OS v6 aims to be a decision support system that adapts to Shapiro's patterns while fostering transparency.

Item	Size	Action
Old backup archive	7.54GB	Deleted
pip cache	~3GB	Cleared
npm cache	~2GB	Cleared
Browser caches	~4GB	Cleared

Location	~Size
/var/lib/docker	16GB
/usr/share/ollama	16GB
HuggingFace cache	20GB

🧠 Cobrain Daemon

Live entity extractor + contradiction flagger. Running in tmux, monitoring chat logs every 60 seconds.

📡 What It Does

Monitors chat logs every 60s
Extracts entities (people, projects, decisions)
Stores in SQLite (cobrain.db)
Audits contradictions against SOUL.md / USER.md / MEMORY.md

🔌 Quick Commands

# Attach to daemon
tmux attach -t cobrain

# Query memory
python3 -c "from cobrain_daemon import recall; print(recall('NOMA'))"

# View DB
sqlite3 /data/telegram_extracted/cobrain.db '.tables'

📁 File Locations

File	Path
Daemon source	`/data/telegram_extracted/cobrain_daemon.py`
SQLite DB	`/data/telegram_extracted/cobrain.db`
Gold Synthesis	`/data/telegram_extracted/GOLD_SYNTHESIS.md`
Brain Synthesis	`/data/telegram_extracted/BRAIN_SYNTHESIS_FINAL.md`