Get the Self-Learning Agent Operator Playbook

Ten sections, every artifact you need to build a self-learning agent system tonight inside Claude Code — without waiting for Anthropic to ship dreaming as a product. The 4-layer memory model from Mahesh Murag's Code-with-Claude SF talk (in-context, in-weights, cache, external — and why layer 4 is where almost all the leverage is). The 4-piece dream loop architecture: CLAUDE.md memory store, .dream/sessions/ logs folder, the /dream slash command, dream-proposal.md review gate. The CLAUDE.md schema with frontmatter — four top-level sections (Decisions, Patterns, Mistakes, Current State) and three metadata fields per entry (date, use_count, confidence) that let the dream curate itself. The complete /dream slash command, every line paste-ready, drop into .claude/commands/ and run it forever. The session-log template that turns every Claude Code run into raw material for the next dream. The four consolidation operations explained with real examples (NEW, UPDATE, DELETE, PROMOTE — the entire API of dreaming). The two-tier review gate ruleset that keeps 90% on auto-apply while routing the 10% high-risk changes to human review. The trigger-mode decision tree (manual, scheduled, idle-triggered) with the operator rule nobody tells you — never dream on fewer than 5 sessions of new data. And the 7 mistakes from running this loop for three weeks, each with the exact fix that prevents it.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.