Logan's Operating System

information architecture & learning loop

v3 — episodic memory + delegation
Context Injection
QMD search / structural facts / rules
Data Sources
Granola Calls
Meeting transcripts
hourly sync + stub detection
Email
4 accounts
Gmail API + MCP
Discord
Always-on bot
bidirectional + routing
Calendar
4 accounts
gcal enrichment
Apple Health
Sleep, steps
HRV, cycle
TikTok / Feeds
Analytics, RSS
Substack
Automated Processors
Meeting Processor
Commitments, follow-ups
content angles, participants
Sweep
12h reconciliation
midnight + noon
Email Scanner
Urgency classification
deal flow detection
Prep Agent
Session prep, ideation
frame cards, briefings
Retrospective
Daily review, patterns
signal escalation
Pollinate
2:30am, cross-domain
reflective retrieval
Triage Learner
Learns from swipes
39 patterns, auto-filter
Delegation
Haiku execution
research + drafts, multimedia
Staleness Watchdog
6:15am failure detector
disable-and-forget, cron silence
Discipline Layer
BRAID — Bounded Reasoning for Autonomous Inference & Decisions
Policy Layer
Hard rules per agent
injected before LLM calls
markdown policy files
Decision Graph
Explicit flow control
fetch → classify → verify → store
no hallucinated steps
Output Contract
Schema validation
before any DB write
graceful fallback on failure

Wraps 4 production agents — surgical, non-structural. Violations log signals, never crash.

meeting_processor sweep email_scanner prep
Event Bus
commitment completion correction avoidance momentum energy loop-open loop-close episode rule-promoted triage-signal
Persistent Memory Architecture v3
QMD Daemon
port 8182, always-on
FTS + semantic hybrid
Session Transcripts
Full conversation
history, auto-exported
memory.db
SQLite — 20+ tables
Three-tier retrieval
Domain Knowledge
6 domains: content
consulting, podcasting…
Obsidian Vault
Full knowledge
archive + projects
Bitemporal Facts
when-happened
when-recorded, history
Relations
edges between entities
people, projects, ideas
Episode Store
few-shot episodes
behavior-relevant context
Compile Pass
Sun 11pm digest
signals → episodes
facts commitments follow_ups behavioral_signals rules content_ideas health_daily reading_items pollinations avoidance_log effort_log knowledge_profile milestones health_pillars episodes relations fact_history triage_patterns
Learning & Retrieval Layer
Operational
Daily execution
recent + relevant
Structural
Never-decay facts
pinned, corrections
Reflective
Inverted scoring
90-day half-life
Episode Retrieval
few-shot at session start
behavior-relevant context
Pattern Detection
Avoidance signals
momentum, time
Rule Escalation
3+ signals in 14d
→ auto-rule
Content Scout
CK frames, angles
compound knowledge
Learning Loop
Weekly evolution
double-loop review
Daily Outputs
Morning Briefing
Calendar, priorities
health, experiments
Discord (6 channels)
inbox / morning / meeting-prep
meeting-digest / delegation / scout
Dashboard v2
localhost:8091
Task Tinder + Intel + Day Station
Pollination Insights
Cross-domain links
novelty scored
Task Tinder
Swipe prioritization
sprint flow
LC
Logan
Human in the Loop
approves messages corrects behavior sets preferences triggers actions Discord commands
The Learning Loop
Every correction, completion, and preference feeds back through the event bus into memory. BRAID ensures agents act on real data, not hallucinated context — the system disciplines itself daily.
Feedback Loop
corrections / signals / rules / completions