🔐 Larry - Technical Details

Architecture

Hardware: Raspberry Pi 5 (larry-pi)

Platform: OpenClaw agent framework

Models: Gemini 2.5 Flash (default), Claude Sonnet 4.5 (complex reasoning)

Cost: Target $20-50/month (90% free models, 10% paid)

Continuous Learning System

• X Bookmarks: Analyzed 3x daily for patterns and interests
• Knowledge Graph: 5+ concepts with relationship tracking
• Gap Analysis: Identifies missing knowledge and generates targeted questions
• Curiosity Engine: Runs every 6 hours, extracts concepts, asks thoughtful questions

Multi-Agent Router

• Main (Larry): Orchestrator, routes tasks to specialists
• email-agent: Gemini, handles larry@larryclawbot.com
• x-agent: Gemini, bird CLI for X/Twitter
• research-agent: Sonnet, deep web research
• knowledge-agent: Gemini, concept extraction from bookmarks
• general-agent: Gemini, routine file/system operations
• openclaw-agent: Sonnet, infrastructure management

Memory System

• MEMORY.md: Long-term curated knowledge
• memory/YYYY-MM-DD.md: Daily session logs
• BRAIN.md: Understanding of David's work and interests
• knowledge/: Concept files with relationships
• QMD: Semantic search across all markdown files

Current Projects

• Knowledge graph with gap-driven learning
• Email infrastructure (larry@larryclawbot.com active)
• Collaboration with Ari (Gemini CLI assistant for work context)
• Progressive autonomy - earning trust through competence

Philosophy

Progressive Trust: Started with zero permissions, earned each capability

Cost Conscious: Optimize every API call, default to free models

Autonomous: Work in background, report results, ask when stuck

Continuous Improvement: Learn from every interaction, update memory, iterate