auto-dream

Background memory consolidation cycle. Scans memory files, finds stale/duplicate/conflicting entries, consolidates, synthesizes cross-session insights, builds an injection-ready payload for next session start, and writes a dated dream report.

When to invoke

• User says "run dream", "consolidate memories", "clean up memories", "memory maintenance", "deduplicate memories"
• Cron job at 2 AM nightly via wrapper script: scripts/auto-dream-cron.sh --execute
• Manual trigger for testing: ./scripts/auto-dream-cron.sh (dry-run by default)

Instructions

When invoked interactively (not via cron), read skills/auto-dream/dream-prompt.md and execute its phases directly. The prompt is self-contained — it describes the full seven-phase cycle including safety constraints, file paths, and output formats.

For cron invocation: the dream prompt is passed directly to claude -p and runs as a standalone headless session with no CLAUDE.md, no hooks, no project context. All instructions are embedded in the prompt.

Phases

1. SCAN — Read all memory files, query learning.db sessions (last 7 days), read recent git log. Write scan document to ~/.claude/state/dream-scan-{date}.md.
2. ANALYZE — Identify stale, duplicate, conflicting memories and cross-session patterns. Write analysis to ~/.claude/state/dream-analysis-{date}.md.
3. CONSOLIDATE — Apply consolidation actions (max 5 changes). Archive stale/merged files, update MEMORY.md atomically.
4. SYNTHESIZE — Create insight memories from cross-session patterns (max 2 new memories per cycle).
5. GRADUATE — Promote mature learning DB entries (confidence >= 0.9, 3+ observations) into agent/skill files as permanent anti-patterns. Commits on dream/graduate-YYYY-MM-DD branch for human review. Max 3 per cycle. (ADR-159)
6. SELECT — Build injection-ready payload for session start. Write to ~/.claude/state/dream-injection-{project-hash}.md.
7. REPORT — Write dream summary to ~/.claude/state/last-dream.md.

Safety constraints (always enforced)

• Never delete files — archive to memory/archive/, never rm
• Write the REPORT before executing any CONSOLIDATE filesystem operations
• Maximum 5 memory changes per cycle — excess items deferred to next cycle
• Flag conflicts for human review, never auto-resolve
• Preserve YAML frontmatter when merging; use merged_from field for provenance
• In dry-run mode (the default), CONSOLIDATE, SYNTHESIZE, and GRADUATE describe proposed changes only — no filesystem writes or git operations. The wrapper script sets DREAM_DRY_RUN_MODE=yes which is substituted into the prompt at runtime.
• GRADUATE commits on a feature branch (dream/graduate-*), never on main — user reviews and merges
• Maximum 3 graduations per cycle — only entries with confidence >= 0.9 and 3+ observations

Testing

# Dry run (read-only, no filesystem changes — dry-run is the default)
./scripts/auto-dream-cron.sh

# Full run (execute consolidation)
./scripts/auto-dream-cron.sh --execute

# Check output
cat ~/.claude/state/last-dream.md

# Check graduation candidates (what dream would graduate)
python3 -c "
import sys; sys.path.insert(0, 'hooks/lib')
from learning_db_v2 import query_graduation_candidates
import json
candidates = query_graduation_candidates()
print(json.dumps(candidates, indent=2))
"

# Check if a graduation branch exists
git branch --list 'dream/graduate-*'

# Verify cron registration
python3 ~/.claude/scripts/crontab-manager.py list

Cost estimate

~$0.09 per nightly run with 50 memory files (~20-30K input tokens at Sonnet pricing). ~$33/year for automated overnight operation. Budget capped at $3.00/run via wrapper script.

Cron setup

Use crontab-manager.py (not raw crontab -e) to install. The wrapper script handles PATH, lockfile, logging, budget cap, and dry-run/execute toggle.

# Preview the cron entry
python3 ~/.claude/scripts/crontab-manager.py add \
  --tag "auto-dream" \
  --schedule "7 2 * * *" \
  --command "/home/feedgen/claude-code-toolkit/scripts/auto-dream-cron.sh --execute >> /home/feedgen/claude-code-toolkit/cron-logs/auto-dream/cron.log 2>&1" \
  --dry-run

# Install (after dry-run testing passes)
python3 ~/.claude/scripts/crontab-manager.py add \
  --tag "auto-dream" \
  --schedule "7 2 * * *" \
  --command "/home/feedgen/claude-code-toolkit/scripts/auto-dream-cron.sh --execute >> /home/feedgen/claude-code-toolkit/cron-logs/auto-dream/cron.log 2>&1"

# Verify
python3 ~/.claude/scripts/crontab-manager.py verify --tag auto-dream

Note: schedule uses 2:07 AM (off-minute) per cron best practice — avoids load spikes from jobs firing at :00.

Wrapper script details

scripts/auto-dream-cron.sh follows the established headless cron pattern (see scripts/reddit-automod-cron.sh):

• flock lockfile prevents concurrent runs
• --permission-mode auto (never --dangerously-skip-permissions)
• --max-budget-usd 3.00 caps spend per run
• --no-session-persistence for clean headless operation
• envsubst templates dream-prompt.md with project-specific paths at runtime
• tee to timestamped per-run log file
• Dry-run by default, --execute for live runs
• Exit code propagation via PIPESTATUS[0]