Jarvis.
A personal AI that actually stays with you.
Not a chatbot. A 24/7 operator running on my gaming laptop: event-sourced memory, macro brain, autonomous work queue, and a command-center UI. Less "how can I help" and more "here's what you missed while you were asleep".
Memory, structured like a brain.
Most "AI assistants" forget you between conversations. Jarvis doesn't. Every layer has a purpose, every source feeds the Cortex, and every session compounds on the last.
Working memory
state.md is the RAM. Current priorities, active trades, unresolved threads. Read first, every boot.
Event log
events.jsonl is the append-only fact stream. Every frontend (Telegram, Claude Code, cron) writes here. The truth.
Boot context
boot_context.py builds a situational awareness blob from time, deadlines, journal, notes, tasks. One pull, full picture.
Decision matrix
decisions.md maps situations to tool chains. No guessing, no tool sprawl. Right tools, every time.
Cortex engine
cortex.py indexes 16 source types. Cross-layer search across sessions, journals, trades, knowledge, scripts.
Long-term memory
User profile, feedback, strategy, references. Compounds across sessions, never overwritten, always evolving.
Macro brain + soul
13 knowledge modules, awareness daemon, anticipation engine, initiative rules, idle work queue. The autonomy layer.
Co-pilot, not chatbot.
Built for a single user (me) over months. Every feature is driven by a real workflow I ran into: trading, study, research, autonomy.
Event-sourced memory
Every action, every trade, every note lands in events.jsonl. Frontends stay in sync because the log is the truth; state files are just snapshots.
Macro brain
13 modules: regime detection (8 states), sentiment model, news mapper, session synthesizer. Institutional-grade analysis on a personal budget.
Self-evolving
Awareness daemon monitors 8 domains. Anticipation engine learns session patterns. Initiative engine fires tiered actions from events. Idle hours become work.
Trading co-pilot
Two-phase workflow: 2 PM Macro Brain + HTF, 4:45 PM LTF + scenarios, 5 PM silent monitor. Pre-session briefs, live price polling, post-session review, all automated.
Command center UI
Next.js frontend: War Room dashboard, brief timeline, Cortex search, Macro Brain, Research runs, System evolution, Activity feed. Live WebSocket updates.
Multi-surface
Same brain, three frontends: Telegram bot (on-the-go), Claude Code terminal (deep work), Next.js UI (cockpit). One event log reconciles all of them.
One view, every signal.
A Next.js command center wired to the brain over WebSockets. Same state everywhere: briefs, trades, research, system health, events.
The stack.
Python brain, Next.js cockpit, Claude at the core, everything glued with MCP and cron. Running self-hosted on a gaming laptop, not the cloud.
Python
Next.js
FastAPI
React
TypeScript
Tailwind
Claude (Anthropic)
MCP
Redis
n8n
Nginx
TradingViewStill evolving.
Jarvis is a year-long project and a personal one, not open source. Want to talk about personal AI, event-sourced memory, or building co-pilots that actually live with you?