Dream

Dream: Memory Consolidation Skill

Invoke: /dream (project) | /dream user (user-level) | /dream all (both)

Reflective pass over Claude Code's auto-memory files. Synthesizes recent learnings into durable, well-organized memories so future sessions orient fast.

Scope Resolution

project:
  memory_dir: .claude/memories/
  index_file: .claude/memories/MEMORY.md
  transcripts_dir: .claude/transcripts/
  index_max_lines: 150

user:
  memory_dir: ~/.claude/memories/
  index_file: ~/.claude/memories/MEMORY.md
  transcripts_dir: ~/.claude/transcripts/
  index_max_lines: 200

Flag behavior:

• /dream → project scope only (default)
• /dream user → user scope only
• /dream all → project first, then user

Phase 1 — Orient

graph LR
  A[ls memory_dir] --> B[Read INDEX_FILE]
  B --> C[Skim topic files]
  C --> D{logs/ or sessions/ exist?}
  D -->|yes| E[Review recent entries]
  D -->|no| F[Proceed to Phase 2]
  E --> F

1. ls the memory directory to see what exists
2. Read ${INDEX_FILE} — understand current index
3. Skim existing topic files to avoid creating duplicates
4. If logs/ or sessions/ subdirs exist, review recent entries

Phase 2 — Gather Recent Signal

Priority order:

1. Daily logs (logs/YYYY/MM/YYYY-MM-DD.md) if present
2. Drifted memories — facts that contradict current codebase/state
3. Transcript search — narrow grep ONLY, never read whole files:

grep -rn "<narrow_term>" ${TRANSCRIPTS_DIR}/ --include="*.jsonl" | tail -50

NEVER-LIE Rule (inherited)

• Only consolidate VERIFIED facts
• If uncertain, grep transcripts to confirm before writing
• Wrong memory = worse than no memory

Phase 3 — Consolidate

For each thing worth remembering, write or update a memory file at the top level of the memory directory. Follow auto-memory conventions from system prompt.

Focus:

• Merge new signal into existing topic files (no near-duplicates)
• Absolute dates — convert "yesterday", "last week", "next Friday" → 2026-03-25
• Delete contradictions — if today's investigation disproves old memory, fix at source
• Context compression — Mermaid for flows, YAML for state, prose for NOTHING

Classification for merge decisions:

actions:
  duplicate: merge into existing file, delete duplicate
  contradiction: keep newer fact, delete older, log change
  stale: delete if >30 days with no transcript reference
  relative_date: convert to absolute ISO date
  verbose: compress to YAML/Mermaid, remove prose
  code_convention: belongs in CLAUDE.md or .claude/rules/, NOT memory
  sensitive: never store tokens/keys/passwords in memory

Phase 4 — Prune & Index

Update ${INDEX_FILE} to stay under ${INDEX_MAX_LINES} lines.

Rules:

• Index = pointers with one-line descriptions, NOT content dumps
• Remove pointers to stale/wrong/superseded memories
• Demote verbose entries: gist in index, detail in topic file
• Add pointers to newly important memories
• Resolve contradictions — if two files disagree, fix the wrong one

Output

Return a structured summary:

dream_summary:
  scope: project|user|all
  date: YYYY-MM-DD
  consolidated: [list of merged files]
  updated: [list of modified files]
  pruned: [list of deleted/removed files]
  unchanged: [list of files kept as-is]
  issues_found:
    duplicates: N
    contradictions: N
    stale: N
    relative_dates: N
    verbose: N
  index_lines_before: N
  index_lines_after: N

If nothing changed (memories already tight), say so.

Integration Notes

• Runs AFTER session work, BEFORE /compact or session kill
• Complements 50% context rule: dream keeps memory lean between sessions
• Does NOT touch CLAUDE.md or .claude/rules/ (Layer 1-3) — memory only
• Safe to run nightly via AUTOLOOP if desired