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.

Signal categories to look for:

Category	Type	Trigger phrases
User corrections	feedback	"no", "don't", "stop doing", "not like that", "actually..."
Confirmed approaches	feedback	"yes exactly", "perfect", silent acceptance of unusual choice
Role / expertise	user	"I'm a ...", "I've been writing X for Y years", "I own the ..."
Project decisions	project	"we're going with", "we decided", deadlines, incidents
External pointers	reference	Dashboard URLs, issue trackers, Slack channels, doc links

What NOT to gather (strict exclusions):

• Code patterns, architecture, file paths, project structure — derivable from reading code.
• Git history or who-changed-what — git log / git blame are authoritative.
• Debugging solutions — the fix is in the code; context is in the commit message.
• Anything already in CLAUDE.md files.
• Ephemeral task state or current conversation context.
• PR lists / activity summaries — save only what was surprising or non-obvious.

---

Phase 3 — Consolidate

Before making any file modifications, show the user a plan:

• List files to be created, updated, merged, or deleted.
• Wait for user approval before proceeding.

After approval, execute:

3a. Repair missing frontmatter

For any topic file flagged in Phase 1, add the required YAML frontmatter:

---
name: {{memory name}}
description: "Consolidates and prunes the memory directory through a 4-phase reflective pass: Orient, Gather, Consolidate, Prune. Run after intensive sessions, when memories feel duplicated, or when MEMORY.md approaches the 200-line limit."
type: {{user | feedback | project | reference}}
---

3b. Merge duplicates and near-duplicates

If two or more files cover the same topic (e.g., feedback_testing_1.md and feedback_testing_2.md), merge them into one canonical file. Delete the originals after merging.

3c. Write / update memories for new signal

For each new signal found in Phase 2, either update an existing file or create a new one.

Memory type conventions:

• feedback — Lead with the rule, then **Why:** (the reason the user gave) and **How to apply:** (when/where this kicks in). Always include Why — it enables edge-case judgement instead of blind rule-following.
• project — Lead with the fact/decision, then **Why:** (motivation, constraint, deadline) and **How to apply:** (how this shapes suggestions). Convert all relative dates to absolute dates (e.g., "Thursday" → "2026-04-10").
• user — Factual profile: role, expertise, preferences, responsibilities.
• reference — Pointer + purpose. One or two lines.

3d. Fix contradictions

If today's investigation proves an old memory wrong, fix it at the source — edit or delete the old file. Do not leave contradicting entries.

---

Phase 4 — Prune and index

Rebuild MEMORY.md so it stays under 200 lines AND under ~25KB.

Rules for the index:

• Each entry is exactly one line, under ~150 characters: - [Title](file.md) — one-line hook
• Never write memory content directly into MEMORY.md — it's an index, not a dump.
• Remove pointers to deleted or superseded files.
• If an index line is over ~200 chars, shorten it — move the detail into the topic file.
• Organize semantically by topic (user → feedback → project → reference), not chronologically.

---

Output

Return a brief summary of what was consolidated, updated, merged, pruned, or repaired. If nothing changed (memories are already tight), say so explicitly.

End with a stats line in this exact format:

Dream complete: X memories total | Y created, Z updated, W merged, V pruned.

(Omit categories with a count of 0.)