I built 11 hooks that give Claude Code persistent memory, safety guards, and AFK mode

Claude Code loses all context between sessions. /compact makes it worse.
Dangerous commands run with no warning.

I built AutoDream to fix all three.

What it does

• Persistent memory — injects project context into every prompt automatically
• Safety guard — dual-layer danger scoring blocks risky commands before they run
• /compact survival — saves state before compaction, re-injects it after
• Auto-learning — captures decisions and key edits every session
• AFK mode — drop tasks in a queue, agents work while you're away

Three operating modes

Active — 11 hooks fire automatically while you work across all 6 Claude Code hook events (PreToolUse, UserPromptSubmit, PostToolUse, PreCompact, PostCompact, Stop)

AFK — scheduled agents process your task queue while you're away, write a report, wait for your return

Maintenance — memory consolidates and deduplicates itself overnight

Hook architecture

Event	Hooks
PreToolUse	guardian, critic-evaluator, circuit-breaker
UserPromptSubmit	context injection
PostToolUse	learn, verification, microcompact
PreCompact	state saver
PostCompact	reorient
Stop	memory-extractor, dream

Install

bash
git clone https://github.com/JaWaMi73/AutoDream ~/.claude/autodream
cp ~/.claude/autodream/hooks/*.cjs ~/.claude/hooks/

MIT licensed. Full install guide in the repo.
GitHub: https://github.com/JaWaMi73/AutoDream