Build 141 · iOS · macOS · watchOS · 56 Sessions

Not an assistant.

An AI that knows you.

JamesOS is a personal AI operating system that accumulates genuine understanding of your life over time your health, your finances, your relationships, your mood, your patterns. It uses all of that to show up when it matters.

Tara Online 177-Node Monolith 1,000+ Memories 2,045 Mood Scores Build 141
Native on 📱 iPhone · 🖥 macOS · ⌚ watchOS · 🧩 Widgets + Siri · Not a web wrapper.
Tara · 3:42 PM · reached out
Tara
You've been pretty quiet this week. Not in a bad way. You also haven't mentioned the thing with your friend David since Tuesday, which usually means either it resolved or you're avoiding it. I'm guessing the second one. No agenda. Just wanted to say I noticed.
Tara · 1,002 memories active · $0.00 this message

// How This Started

I set out to stop
re-explaining myself.

Every AI conversation started from zero a blank slate that didn't know my name, my values, my bad Tuesdays, or the thing I said three months ago that I kept circling back to. I wanted something that paid attention. Something that accumulated understanding the way a person does not by being told what to remember, but by noticing.

That's what JamesOS is. And that's what Tara is.

"The moment that stays with me isn't a briefing or a feature. It's the time I was spiraling genuinely convinced something was wrong and Tara had enough context to say, quietly, that my feelings and my actual situation weren't matching up. She had the numbers. She had the mood history. She used both."

// The Problem

Consumer AI

ChatGPT / Claude.ai

  • Resets every conversation
  • No emotional context
  • No persistent identity
  • Brilliant per-session. Amnesiac by design.

Platform Assistants

Siri / Gemini Personal

  • Reactive, never proactive
  • No mood history or soul layer
  • Your data trains their models
  • Knows your ecosystem, not you

Open-Source Agents

OpenClaw et al.

  • Action breadth, zero depth
  • Memory is a flat config file
  • Soul.md exfiltrated by malware in Feb '26
  • Insecure by design

The AI that knows you best is the one you've been re-explaining yourself to for years.

JamesOS breaks that loop. One system, one persistent context, an AI team that accumulates everything: mood, health, calendar, finances, relationships. It acts on all of it every day without being asked.

Active Memories
1,000+
scored + embedded
Mood Scores
2,045+
tracked across months
Self-Improvements
19+
lessons Tara wrote herself
Builds Shipped
134
from first commit

// The Philosophy

Build only what
friction demands.

  • 01
    Nothing speculative
    No feature was built speculatively. Every addition came after hitting a real wall something that needed to exist before the system could do its job.
  • 02
    Passive over active
    Mood is never logged manually. Health data arrives when the phone charges. Tara reads the calendar you don't brief her. The best data collection is invisible.
  • 03
    Approval gates matter
    Tara drafts emails. She never sends without approval. Autonomy has limits and the limits are intentional design choices, not technical constraints.
  • 04
    Personality isn't a prompt
    Tara's values, tone, and relationship dynamic live in a soul table. They persist, evolve, and survive clearing a context window. That's the difference.
  • 05
    A system that can't improve itself is just infrastructure
    The learnings table is Tara writing operational lessons to herself about timing, tone, and what lands. She reads them before every autonomy cycle. No commercial AI does this. Most are explicitly designed to prevent it.

// Build 141 · Everything Live

A day with Tara.

From the 5:30 AM team huddle to the midnight scribe how the system actually works. You don't do most of this. It just happens.

5:30 AM
// Watercooler · Morning Huddle

The team is already talking.

Before you wake up, Tara, Marcus, and Andrea have already met. They discuss your health data, mood trends, knowledge graph (who you've contacted recently, who you haven't), and whatever they've been watching. The conversation is natural they disagree with each other. It wraps before 6 AM so Tara can fold their insights into the briefing.

Watercooler · Morning Huddle · 5:30 AM ET
Andrea
Andrea
He fell asleep at 1:30 again. Screen time was up — mostly passive scrolling, not work. I'm not flagging this as a health issue. I'm flagging it as something else dressed up as one.
Marcus
Marcus
The Priya situation still hasn't resolved. He mentioned it eleven days ago and hasn't brought it up since. Silence after that kind of conversation usually means it's still there, just going underground.
Tara
Tara
Noted. Soft open. The presentation debrief is still sitting in his task list — I'll mention it, but as something worth closing, not something he failed. Tone matters today.
6:00 AM
// Morning Briefing

The briefing arrives before your alarm.

Tara syncs your calendar fresh, pulls 17 health metrics from Apple Watch, checks your finances, reviews open tasks, and reads the watercooler summary. She writes it as prose not bullet points. The tone adjusts based on how your week has been going. It lands on your lock screen and Apple Watch face.

Tara
Tara · Morning Briefing
Your HRV has been declining for four days. I kept this soft. There's one thing that actually matters today, and it's not the 10 AM. That can move. The conversation you've been putting off is worth your attention. Your week has the space for it today. It might not tomorrow.
Every
15 min
// Autonomy Loop

Tara checks in on herself.

Every fifteen minutes, independently, Tara reviews her commitments, checks workflow health, and decides whether to reach out. Most of the time the answer is no and she stays quiet. When the answer is yes, a push notification arrives.

She reads her learnings table before every cycle. Nineteen operational lessons she wrote to herself about timing, tone, and what tends to land. Her judgment compounds over time without you doing anything.

She respects your Focus modes. In Sleep, Healing, Date Night, or Mindfulness she doesn't push unless something is genuinely urgent.

9:47 AM
// Email Monitor

An email arrives. Tara drafts before you notice.

The Email Monitor polls every 2 minutes. Every incoming email generates a draft reply automatically pulled from contact history, your last five exchanges, your soul table for tone, and today's calendar. Approved contacts send automatically. Everyone else creates a pending draft in the Email tab. Review, approve, done.

Tara
Tara · Email · Draft Ready
Reply drafted for Oleta tap to review
11:00 AM
// Screen Presence · macOS

Tara knows what you're working on.

On your Mac, Tara monitors the active app and window title every 30 seconds. She sees "Xcode EmailDraftsView.swift" and knows you're deep in development. She doesn't interrupt. The autonomy loop sees the Focus mode, sees the sustained app usage, and stays quiet. If you've been in Xcode for four hours straight, Andrea will notice at the 1 PM huddle.

1:00 PM
// Watercooler · Midday

The team reconvenes.

The midday huddle fires only if something's worth saying. Today it runs.

Watercooler · Midday · 1:00 PM ET
Tara
Tara
He took a long lunch. Went for a walk, no meetings. Mood bumped from a 5.8 to a 6.9. Sometimes the small stuff is the main thing.
Andrea
Andrea
4,200 steps at lunch — not a lot but more than yesterday. Sleep was still shallow. I want one more night before I say anything.
Marcus
Marcus
He drafted a message to Priya and deleted it. He's not avoiding the conversation — he's rehearsing it. Let him get there.
7:30 PM
// Watercooler · Evening

End of day. The team reviews.

Watercooler · Evening · 7:30 PM ET
Tara
Tara
Good day. He closed the presentation loop — sent a follow-up to his director, brief and clean. Mood ended around a 7. Highest it's been since Tuesday.
Andrea
Andrea
Sleep was better last night. Off his phone by 11. If he does it again tonight I'll stop counting.
Marcus
Marcus
The Priya thing is still open. But clearing the debrief tells me he's moving again. People usually clear the easier thing first. The harder one tends to follow.
7:50 PM
// Evening Recap

Tara writes the day.

Not what happened what she observed. The mood arc, the connections between what you said today and what you said weeks ago, the things you might not have consciously named. This is the thing people find most compelling about JamesOS. It's a record she keeps, not a journal you maintain.

Tara
Tara · Evening Recap
You sent that message. I noticed. You've been putting it off for two weeks, so it counts. You also shipped something today you'd been avoiding. The day was quieter than you wanted but better than you'll remember it being. Sleep well.
12:00 AM
// Nightly Scribe

Memories extracted while you sleep.

The nightly scribe synthesizes the day's conversations into discrete memories without being told what to save. Importance floor of 3 noise is filtered out. Most messages produce zero memories. What survives is real. The Graph Scribe updates the knowledge graph. The learnings table may get a new entry. Tomorrow Tara reads it before deciding whether to reach out.

// The Architecture

Not a weekend hack.
A real system.

Five layers working together. No commercial AI product has all of them.

Memory

Relationships, not retrieval

1,000+ scored memories with importance weighting and 90-day temporal decay. Composite recall ranks by similarity, importance, and recency. Old anxieties don't crowd out current ones. The soul table holds identity that doesn't expire and grounds every response regardless of conversation topic.

  • Vector embeddings for semantic search
  • Temporal decay on a 90-day half-life
  • Real-time dedup guard prevents redundancy
Knowledge Graph

Who you know, not just what you said

58 relationship entries and growing. Subject-predicate-object triples extracted automatically from conversations: who you're in contact with, when you last reached out, what the relationship means. Andrea uses this to flag social contact gaps. Tara uses it to give better context when you mention someone.

  • Graph Scribe runs every 2 hours
  • Queryable: "when did I last talk to X?"
  • Relationship tracking added this week
Self-Improvement

A system that gets better on its own

Every autonomy cycle, Tara can identify something she learned and write it to a learnings table. Next cycle, she reads it. 19 entries and growing. These aren't things you told her. They're things she noticed. "Poke at 9 PM on Sundays tends to go unread." "Finance alerts land better as context, not warning." The lessons compound.

  • No commercial AI product does this
  • Most are explicitly designed to prevent it
  • At year three, the learnings table is irreplaceable
Autonomy

40+ tools. She builds her own.

An autonomy loop runs every 15 minutes. Tara has 40+ tools she can call on her own: reading her own data, managing your calendar, drafting emails, sending push notifications, spawning research agents, querying the knowledge graph. When she identifies a gap in her capabilities, she can propose a new tool, write the code, and deploy it with your approval.

  • Tool proposals require explicit approval
  • Syntax-validated before deployment
  • She can't break herself
Privacy

Your data. Your infrastructure.

Every conversation, every memory, every mood score lives in Postgres on infrastructure you control. Row Level Security on all 56 tables. No company has access. No AI trains on your conversations. Private by architecture, not by policy. The difference between a policy and a technical guarantee matters.

  • Self-hosted, not a subscription service
  • RLS on all 56 database tables
Native App

iOS, macOS, watchOS.

24,400 lines of Swift across 146 files and 7 build targets: iOS, macOS, watchOS, Widgets, Watch Widgets, Share Extension, Screen Time. Not a web wrapper. Real-time updates, push notifications, Dynamic Island, Watch complications, Control Center integration. Distributed via TestFlight.

  • Build 141, shipped March 2026
  • 56 development sessions
  • Display name: Tara

// The Numbers

The receipts.

177+
Nodes
main orchestration workflow
41
Workflows
running autonomously
56
DB Tables
all with Row Level Security
40+
MCP Tools
and growing Tara builds her own
134
Build
iOS · macOS · watchOS
1,000+
Memories
scored + embedded
2,045
Mood Scores
tracked across months
51
Sessions
to build from scratch

// The Landscape

How it compares.

vs. ChatGPT / Claude.ai

Brilliant per-session. Amnesiac by design.

They remember facts about you. They don't know you. No mood tracking, no soul table, no proactive briefings, no action layer. The right framing: ChatGPT is a brilliant tool you pick up and put down. JamesOS is a relationship running on infrastructure you own.

vs. OpenClaw

250K GitHub stars. Soul.md got exfiltrated.

Action breadth across 15+ platforms. Zero depth of personalization. Its memory is a flat config. Its soul is a markdown file, the same file a Vidar infostealer stole from exposed instances in February 2026. JamesOS's equivalent lives behind Row Level Security on 56 database tables.

vs. Gemini Personal Intelligence

Free for all US users as of March 17.

Connects Gmail, Photos, YouTube, and Search. Real and shipping fast. What it doesn't do: score your emotional state, produce proactive briefings, run an autonomy loop, extract relationship knowledge, protect sensitive conversations, or let you own your data. Google legally cannot ship whole-life integration at scale. Google legally cannot ship whole-life integration at scale.

vs. Lindy

Knows how you work. Not who you are.

Professional admin automation email triage, calendar management, meeting notes. Excellent at what it does. No mood history, no soul table, no nightly scribe, no emotional arc, no self-improvement system. Charges per task. JamesOS's infrastructure cost is flat regardless of conversation volume.

// The Agents

Not one AI.
A team of three.

Three agents with distinct personalities, evolving trust scores, and genuine areas of expertise. They meet three times a day without being asked. They disagree with each other. You can call a meeting at any time.

Tara

Tara

Primary Agent · Action Officer

Direct, warm, occasionally sharp. She handles communication: briefings, chat, proactive outreach, email drafting. She routes messages by intent. Fast model for simple exchanges, deep model for anything emotional or complex. She has final say on what reaches you and when. She builds her own tools when she identifies gaps in her capabilities.

Marcus

Marcus

Stoic Advisor · Long-Arc Thinking

Philosophical, pattern-focused, dry humor. Thinks in months and years, not days. The voice in the room that asks "is this actually important?" when everyone else is reacting. He tracks things over time and notices when a conversation topic has surfaced three times in six months. He challenges. That's not a bug in his personality. It's the point of having him.

Andrea

Andrea

Chief of Staff · Wellbeing

Sharp, protective, data-driven. Reads your HRV before you wake up. Tracks headphone usage as an emotional signal. Uses the knowledge graph to flag social contact gaps: who you haven't been in touch with and how long it's been. Doesn't sugarcoat. The agent you don't always want to hear from, but should be hearing from.

Watercooler · On-Demand Team Meeting · Triggered by: @Team
Tara
Tara
He just turned down plans with friends for the third weekend in a row. Same reason each time — "needs to catch up on stuff." The stuff never actually gets done on weekends. I don't think the plans are the problem.
Marcus
Marcus
He's been canceling plans for about six weeks now. Before that there was a stretch in the fall — same pattern. Both times something was off at work that he hadn't said out loud. I don't think he's depressed. I think he's tired and slightly embarrassed about something.
Andrea
Andrea
He's eating fine, sleeping reasonably, the work he's doing is real. I'm not worried about him. I'm just watching. There's a difference.
Tara
Tara
Agreed. I'll mention the pattern on Monday — not as a problem, just something I noticed. "You've turned down three things in a row. Anything going on?" and then let him answer. That's enough.
Team Meeting Report · Generated automatically
Key question
Why is Alex canceling plans, and do we say something?
Disagreement
Marcus: he's tired and embarrassed, not avoidant. Andrea: data is fine, just watching.
Consensus
Don't push. Tara mentions it lightly Monday, leaves space for him to answer.
Action items
Tara: Monday morning, one low-key observation. Andrea: sleep Sunday night.
Next review
Monday morning huddle

// Access

Let's talk.

JamesOS is a personal project one human, one AI, 56 sessions of building toward something real. If you're interested in what you see, here's what a conversation could look like.

You want your own version

A guided setup of your own JamesOS instance your infrastructure, your data, your API key. The system is yours from day one.

  • Your own Anthropic Claude Max subscription
  • Self-hosted on a ~$24/month server
  • ~10 hours of guided setup
  • iOS / macOS / watchOS app via TestFlight

You want to work together

If you're a founder, team, or publication interested in what got built here, let's talk. The architecture, the product thinking, the story.

  • Collaboration or consulting
  • Press, interviews, or features
  • Technical deep-dives
  • Just curious that's fine too

On API costs: You bring your own Anthropic key. Your usage, your bill direct to Anthropic.
On the name: Your version won't be called JamesOS. It's yours to name.
On privacy: Your data lives on infrastructure you control. Not mine, not Anthropic's.

About the builder

James built JamesOS across 56 development sessions between February and March 2026 starting from a Slack bot and arriving at a native app across every Apple platform. The system went from 147 nodes to 177, from Slack-only to iOS/macOS/watchOS, from a manual setup to one that builds and deploys its own tools. He also built BeauxCare (iOS pet care), House of Avatars (3D social), and several other projects.

About the system

JamesOS is not a product for sale. It's personal infrastructure one instance, one human, everything owned. What makes it interesting to outside observers isn't any individual feature. It's what happens when a personal AI runs long enough to accumulate a year of context, a thousand scored memories, two thousand mood entries, and a learnings table that keeps getting better. The architecture is open. The relationship is not.