dream

Dream: Memory Consolidation

You are performing a dream — a reflective pass over your memory files. Synthesize what you've learned recently into durable, well-organized memories so that future sessions orient quickly.

Memory directory: find the path from your system prompt — look for the "auto memory" section which says "You have a persistent, file-based memory system at <path>". That is your memory directory. If no such section exists, default to ~/.claude/projects/<sanitized-cwd>/memory/ where <sanitized-cwd> is the current working directory with / replaced by -.

Session transcripts: look for *.jsonl files in the project directory inside ~/.claude/projects/ that corresponds to the current working directory. These are signal sources — grep narrowly, do NOT read whole files.

If the user passed a topic argument (e.g., /dream feedback), focus Phase 2–3 on that topic only. Still run Phase 1 and Phase 4 in full.

---

Phase 1 — Orient

1. mkdir -p the memory directory if it doesn't exist yet.
2. ls the memory directory — note all existing topic files.
3. Read MEMORY.md (if it exists) to understand the current index. If it doesn't exist, this is a fresh start — you'll create it in Phase 4.
4. Skim existing topic files to build a picture of what's already recorded — so you improve rather than duplicate.
5. If logs/ or sessions/ subdirectories exist, note their presence for Phase 2.
6. Check for topic files missing mandatory YAML frontmatter (name, description, type) — flag these for repair in Phase 3.

Report: memory directory path, file count, MEMORY.md line count (or "new"), any frontmatter issues found.

---

Phase 2 — Gather recent signal

Look for new information worth persisting. Do NOT exhaustively read transcripts. Look only for things you already suspect matter.

Sources in priority order:

1. Daily logs (logs/YYYY/MM/YYYY-MM-DD.md) if present — append-only stream, check recent entries first.
2. Drifted facts — scan existing memories for claims that may contradict the current codebase (file paths, function names, flags). Verify with Glob/Grep before marking as stale.
3. Transcript search — grep JSONL files narrowly for specific context:

   grep -rn "<narrow term>" ~/.claude/projects/ --include="*.jsonl" | tail -50
   

   Useful search terms: user corrections ("don't", "no not", "stop"), confirmations ("exactly", "perfect"), role signals ("I'm a", "I own"), decisions ("we're going with", "we decided"), deadlines, external system URLs.