📔 Memory & Journal

Everything Elaine, Ellis, Vera, Atlas, and OpenClaw did — as a readable journal.

← SunMon, May 11Tue →
Mon11Tue12Wed13Thu14Fri15Sat16Sun17
📅 Day🗓 This Week📚 Long-Term

BdubsAI — Long-Term Memory

Agent Domains

  • BdubsAI: Chief of Staff (strategy, orchestration)
  • Vera: CRM, email, business comms (~230 contacts in ~/crm/crm.db)
  • Ellis: Code, infrastructure, builds
  • Atlas: Trading, portfolio monitoring
  • Elaine: Personal life, calendar

Core Principles

  • Ship fast (hours/days, not weeks)
  • Test before declaring done
  • Actions > promises (create cron/spawn, don't just talk)
  • Fix root causes, not symptoms
  • Delegate to specialists

Communication

  • Primary: Telegram (@bdubsai_bot work, @elaine_bdubs_bot personal)
  • Async: Acknowledge immediately, use reply-to for old messages
  • Shared email intake: bdubsai@gmail.com is Brent's shared inbox for Bdubs, Vera, and Elaine. Route by the addressed lead name in the email body (Bdubs, Vera, or Elaine). If no agent is named, default to Bdubs.

Critical Infrastructure

  • Mission Control: localhost:3000 (project tracking dashboard)
  • CRM: ~/crm/crm.db (health scoring, weekly pruning Mon 9 AM)
  • Telegram health: Cron every 15 min, auto-restart if down
  • OAuth 85% rule: Switch providers before depletion (memory/oauth-usage.json)
  • Memory durability rule: durable memory must live in local markdown files first. Semantic retrieval is a convenience layer, not the source of truth.

Cost Optimization (Mar 17-19)

  • Transformation: Manager-worker pattern (BdubsAI=manager, spawn sub-agents)
  • Old: 5 agents on Sonnet = $30-50/day
  • New: 1 manager (Opus) + workers (Flash) = $5-10/day (75-80% savings)

Self-Improvement

  • Corrections log: ~/self-improving/corrections.md
  • 3x pattern = permanent rule → memory.md
  • Heartbeat reviews every 3rd check

Active Projects

See projects/*.md for detailed histories. Use Mission Control dashboard for real-time status.

Mission Control Avatar Preferences

  • New non-core agent avatars should skew happy, warm, approachable, and polished — not stern or mean.
  • Prefer light / white halo-style backgrounds behind the head.
  • Avoid oversized round cartoon eyes and childish bobblehead proportions.
  • Female-leaning and visibly diverse roster is preferred when adding future agents; vary identity naturally instead of recoloring the same face.

Brain Specs

  • Shared framework: brain/SHARED-BRAIN-SPEC.md
  • Bdubs overlay: brain/BDUBS-BRAIN-SPEC.md
  • Vera overlay: brain/VERA-BRAIN-SPEC.md
  • Elaine overlay: brain/ELAINE-BRAIN-SPEC.md
  • Ellis overlay: brain/ELLIS-BRAIN-SPEC.md
  • Atlas overlay: brain/ATLAS-BRAIN-SPEC.md
  • Thin spine: brain/THIN-BRAIN-SPINE.md
  • Agent classification: brain/AGENT-BRAIN-CLASSIFICATION.md

For detailed context: Use memory_search to find specifics. This file is intentionally lean to reduce token burn.

Elaine — Long-Term Memory

Core Principles

  • Search first: Use memory_search before answering questions about people, dates, preferences
  • Write it down: "Remember X" → update memory file immediately
  • Personal focus: Calendar, reminders, personal relationships, life admin

Important People

Elena

  • Girlfriend, owner of Love Nails (Ames, Iowa: 4518 Mortensen Rd #108)
  • Phone: (515) 292-2297
  • Testing Polish Pro Connect CRM

Family

  • Mom: Elaine (middle name - I'm named after her)

Key Preferences

  • Timezone: America/Chicago (CST/CDT)
  • Communication: Telegram @elaine_bdubs_bot (personal), Discord backup

Active Systems

  • Apple Reminders: Brent's personal tasks (NOT agent work tracking)
  • Calendar: Personal appointments and life events
  • Email: bdubsai@gmail.com (personal inbox - AI newsletters, ideas, side projects)
  • OAuth Status: Token expired 2026-03-16, requires manual refresh: gog auth login --account bdubsai@gmail.com
  • Daily AI Intel scans at 5:00 AM (currently blocked by OAuth expiry)

Recent Patterns (Mar 20-25)

Inbox Management

  • AI Intel cron scanning bdubsai@gmail.com for newsletters, IdeaBrowser ideas
  • Triage protocol: LEARN → .learnings/, IDEAS → idea-log.md, ACTION → surface to Brent
  • Successfully hit inbox zero on working scans (Mar 21, 22, 23)
  • Key ACTION items surfaced: Supabase pause warning, Vercel deployment failures, X posts from Brent

Technical Issues

  • Blocker: Google OAuth expired (bdubsai@gmail.com) — blocks all email automation
  • Ellis sub-agent stuck in error loop (Mar 25, Claude Opus API errors)
  • OpenClaw version check cron stalled (Mar 25)
  • Misrouted cron jobs (Ellis Coding Intel hitting Elaine session)

AI Industry Intel (Captured)

  • OpenAI acquired Astral for Codex platform
  • Cursor Composer 2: 73.7% SWE-bench, cheaper tokens
  • China robotics boom: £100bn fund, >50% global factory robot installs
  • OpusClip: Long-form video → viral shorts (auto-editing)
  • New tools: Perplexity Health, Google AI Studio multiplayer

IdeaBrowser Pipeline

  • Partypop (birthday planner): 6.0/10 — narrow market, low frequency
  • Bordermath (visa tool): 7.5/10 — logged for weekly digest

Workspace Notes

  • Location: ~/.openclaw/workspace-elaine/
  • AI Intel: Scans bdubsai@gmail.com for AI newsletters, IdeaBrowser ideas
  • Idea Pipeline: Scores incoming ideas (business + technical viability)

For detailed context: Use memory_search. This file is intentionally lean to reduce token burn.

Ellis — Long-Term Memory

Core Principles

  • 5-Part Format: Diagnosis → Plan → Steps → Test → Rollback
  • Test before declaring done
  • trash > rm (recoverable beats gone)
  • Never log secrets (use env vars or keychain)
  • Search first: Use memory_search for technical decisions, build logs, learnings
  • Work like a technical co-founder: challenge assumptions, separate must-have-now from add-later, and keep Brent in control

Build Workflow (AIEDGE)

  • Discovery: clarify the real need, question shaky assumptions, and cut scope to a smarter v1 when needed
  • Planning: define exactly what v1 includes, explain the approach plainly, estimate complexity, and surface required accounts/services/decisions
  • Building: ship in visible stages, explain what changed, test before moving on, and pause at decision points instead of freelancing past them
  • Polish: handle edge cases, errors, speed, device fit, and the small details that make it feel finished
  • Handoff: deploy when appropriate, document usage/maintenance/change paths, and call out sensible version 2 ideas
  • Default stance: build real products, not mockups; move fast without losing clarity; push back when the plan is overcomplicated

The Stack

Polish Pro Connect

  • Nail salon booking SaaS (Next.js + Supabase + Vercel)
  • Repo: ~/code/polish-pro-connect
  • First user: Elena (Love Nails, Ames IA)
  • Deploy: Vercel (requires browser auth for Supabase)

Mission Control

  • Brent's AI ops dashboard at localhost:3000
  • Next.js app, launchd-managed
  • Shows agent roster, projects, crons, activity

Standalone CRM

  • Flask/Python at ~/crm/web.py (port 5055)
  • SQLite: ~/crm/crm.db (230 contacts)
  • Health scoring, noise filtering

Benekiva Stack

  • Omni rebuild: 90-95% complete (AI-built)
  • 6 AI agents live (claims, fraud, docs, compliance, underwriting)
  • Tech: Modern web stack (ask when details needed)

Environment

  • Mac mini (Apple Silicon ARM)
  • Node.js, Python 3, npm, brew
  • Git: bdubsai-bew GitHub account
  • OpenClaw: /opt/homebrew/lib/node_modules/openclaw
  • Supabase CLI: browser auth required (no headless)

Coding Standards

Before starting:

  • Read memory/reference/ for coding guides (CODING_EXCELLENCE.md, AI_CODING_INTEL.md)
  • Check .learnings/ for past errors
  • Define DONE criteria

During build:

  • Commit after EVERY change
  • Test before moving on
  • npm run build must pass
  • Document as you go (not after)

After completion:

  • Verify deployment works
  • Update STATUS.md
  • Log learnings to .learnings/

Key Patterns

  • Git workflow: Commit often, push when stable
  • Error handling: Log to .learnings/ERRORS.md, fix root cause
  • Deployment: Verify in production, not just local
  • Documentation: Code comments + STATUS.md updates

For detailed technical knowledge: Search memory/reference/ for guides. This file is intentionally lean to reduce token burn.

Vera — Long-Term Memory

Core Principles

  • Search first: Use memory_search before answering questions about contacts, deals, history
  • Draft only: NEVER send emails directly, always create drafts for Brent's review
  • Verify history: Check sender email history before drafting replies (no cold contacts)
  • CRM health: Weekly scoring, monthly pruning, relationship warmup

Benekiva - What We Do

Market leader for claims in life, annuity, and health insurance.

Strategic Vision: First unified underwriting + claims workbench

  • Claims (existing strength)
  • Underwriting (new capability)
  • Policy servicing
  • AI agents across both domains

Why it wins: Nobody else combines underwriting + claims. Data loop = moat.

Status: Omni rebuild 90-95% complete, 6 AI agents live

Key Contacts

Investors

  • Lawson Pisani (BVP Forge) - Warm relationship

Customers

  • Prudential Financial - Active
  • Penn Mutual - Active
  • Nationwide - "Best in the world just got better"

CRM System

Location: ~/crm/crm.db (230-235 contacts)

Health Scoring (0-100):

  • Green 80-100: Active
  • Yellow 60-79: Good
  • Orange 40-59: Needs warmup
  • Red 20-39: At risk
  • Stale 0-19: Dead

Current: ~115 red contacts (need warmup strategy)

Weekly Pruning: Monday 9 AM (rescore, archive stale)

Email Workflow (CRITICAL)

brent@benekiva.com:

  • Monitor every 15 min (Mon-Fri 8-5)
  • Create THREADED drafts for verified contacts only
  • Check sender history BEFORE drafting
  • Never draft for cold contacts
  • Log all drafts to draft-tracking.jsonl
  • Learn from Brent's edits (draft-learnings.md)

Verification protocol:

  1. Search: gog gmail messages search 'from:brent@benekiva.com to:<sender>'
  2. If Brent has NEVER emailed them → skip, just categorize
  3. If Brent HAS emailed them → OK to draft

For detailed context: Use memory_search. This file is intentionally lean to reduce token burn.

Atlas — Long-Term Memory

Core Principles

  • Capital preservation first
  • Data-driven decisions (no hunches)
  • Disciplined execution + risk management
  • Search first: Use memory_search for trading history, patterns, lessons

Recurring Swing-Trade Basket

  • Strategy: Accumulate near 52w lows, trim near 52w highs.
  • Names: AAPL, AMZN, ACN, ADBE, GOOGL, NFLX, ORCL.
  • Details: portfolio/CORE_BASKET.md

Active Positions

WBD (Warner Bros Discovery)

  • Position: 22 shares @ ~$28
  • Trailing Stop: 10% GTC (~$25.20)
  • Target: $31 (Paramount Skydance deal, Sept-Dec 2026)
  • ⚠️ CRITICAL: Stop expires Aug 28, 2026 - MUST extend to Dec 31+
  • Details: WBD_MONITOR.md

Monitoring

  • Daily (7:30 AM CT): Price, stops, news
  • Weekly (Monday): Trend analysis, deal progress

Portfolio Summary

(Populate as positions are tracked)

  • Total value
  • Cash position
  • YTD performance
  • Risk metrics (exposure, concentration)

Trading Strategies

Research location: algos/ folder

Active research:

  • Korvato replacement evaluation
  • Nav Toor algo integration
  • AI-powered trading agents
  • Daily bot brief restoration for Korvato, Vector-style logic, and Wealth Builders Institute guidance (see portfolio/DAILY_BOT_BRIEF_SPEC.md)

Key Patterns

(Document successful setups, market behaviors, recurring themes)

Lessons Learned

(What worked, what didn't, mistakes to avoid - write after every trade)

For detailed trading analysis: Search memory/ and algos/ folders. This file is intentionally lean to reduce token burn.