dream

name	dream
description	Autonomous nightly memory consolidation for one project.
user-invocable	true

Dream — memory consolidation

Consolidates and deduplicates memory for one project using mem0 classifier + Park reflection. Claude does all reasoning inline; scripts handle only file I/O and git.

Invocation

/dream <project> [--dry-run]
/dream <project> --rollback <commit-hash>

<project> is the directory name under ~/.claude/projects/ (e.g. -home-haduong--claude).

Steps

Rollback mode

If --rollback <hash> is present, run:

python3 ~/.claude/skills/dream/commit.py rollback <hash>

Then stop.

Consolidation

1. Read the memory index.

python3 ~/.claude/skills/dream/read-index.py <project>

Output is JSON: {project, memory_dir, entries[]} where each entry has filename, title, desc, content, path.

If entries is empty, log "No memory entries found" and stop.

2. Classify each entry.

For each entry, read its content and the content of semantically related entries (look for keyword overlap in titles and descriptions). Then decide:

• NOOP — already accurate and not redundant with another entry.
• ADD — genuinely new information with no semantic equivalent (rare; most entries were already classified when written).
• UPDATE — should be merged into or superseded by another entry.
• DELETE — contradicted by a more recent entry, or fully stale.

Rules:

• Preserve entries that document evolution ("use vim" → "use emacs"): keep both unless one is explicitly obsolete.
• Only DELETE if the entry is genuinely misleading or already captured elsewhere.
• Log each decision with one line of reasoning.

3. Reflect on survivors.

From entries classified NOOP, ADD, or UPDATE: extract 5 high-level insights that capture patterns, recurring themes, or important lessons. Each insight: 1–2 sentences. These go into the regenerated index as context above the entry list.

4. Report.

Print a summary table:

• Each filename → decision + one-line reason
• The 5 extracted insights
• Counts: N entries before → M after

If --dry-run: stop here. Do not write anything.

5. Apply deletions.

For each DELETE entry, overwrite its file with a tombstone:

# DELETED <ISO timestamp>: <title>
# Reason: <one-line reason>
# Original content preserved in git history.

Use the Edit or Write tool. Never use rm.

6. Rewrite MEMORY.md.

Write a new MEMORY.md at <memory_dir>/MEMORY.md with this structure:

## Key insights

- <insight 1>
- <insight 2>
...

## Entries

- [<title>](<filename>) — <desc>
...

Include only surviving entries (NOOP, ADD, UPDATE). Keep the index under 200 lines.

7. Commit.

python3 ~/.claude/skills/dream/commit.py commit <project> <n_before> <n_after>

Schedule recipe

One entry per project:

0 2 * * * /dream <project>

To inspect consolidation history:

git log --grep='^dream: consolidate' --oneline

v2 roadmap

• Harness-level memory tier + earned promotion (ticket 0163)
• Importance-weighted trigger (SCM pattern)
• Temporal-hierarchical reflection (TiMem) for multi-session horizons
• Link updates after UPDATE/DELETE passes (A-MEM)