Auto Dream — Memory Consolidation#

Run a targeted memory consolidation pass. The goal is not to archive everything; it is to keep future sessions accurate, small, and useful.

Use When#

• Memory or context feels stale, duplicated, contradictory, or too large
• A long coding/research/ops session produced durable facts or decisions
• Daily reflection completed and durable items need promotion
• Before handing off work to another agent/session
• The user asks to dream, consolidate memory, clean memory, or prune context

Core Rules#

• Merge before creating: update existing topic files instead of making duplicates
• Preserve durable facts, decisions, preferences, project conventions, and recurring tool gotchas
• Remove or downgrade task progress, stale todos, solved bugs, and noisy session logs
• Convert relative dates like “yesterday” or “last week” to absolute dates
• Never store secrets, tokens, customer data, private raw transcripts, or full command output
• Make the pass idempotent: running twice should not create extra churn

Four-Phase Workflow#

1. Orient#

Build a quick map of existing memory before editing.

Check, when present:

• MEMORY.md or workspace memory index
• memory/YYYY-MM-DD.md for today and yesterday
• project/topic memory files relevant to the current work
• memory/morning-briefing.md only as an index, not as source of truth

Do not exhaustively read transcripts. Search targeted keywords only when needed.

2. Gather Signal#

Collect only durable signal:

• user corrections and stable preferences
• project decisions and architecture constraints
• solved recurring errors and tool quirks
• important security, auth, deployment, billing, or data-handling rules
• new source-of-truth paths or ownership changes

Discard:

• routine progress updates
• “we did X today” logs with no future value
• temporary todos
• duplicate phrasing of facts already stored
• private/customer data examples

3. Merge#

For every useful item:

• update the closest existing file/section
• resolve contradictions by keeping the newer verified fact and removing stale wording
• compress to one clear line when possible
• move details out of the global index into topic/project files
• mark resolved items only if that state will matter later; otherwise remove them

MEMORY.md should stay an index, not a transcript dump:

• target: under ~200 lines and ~25 KB
• one fact per line when practical
• link or point to detail files instead of duplicating detail

4. Prune#

Remove or shorten:

• completed one-off tasks
• stale debugging notes older than 30 days unless recurring
• duplicate project facts
• obsolete paths, credentials notes, setup commands, or ownership claims
• long entries that can become concise references

Output#

Return a short summary:

text

🌙 Dream abgeschlossen
Merged: X durable facts into existing memory
Pruned: X stale/duplicate entries
Fixed: X contradictions
Memory size: before → after
Open questions: none | list

If nothing changed, say: Memories are already tight — nothing to do.

Cron Trigger Heuristics#

Run automatically when:

• today's daily memory exceeds ~100 lines
• global memory approaches its size/line budget
• a session longer than 30 minutes changed durable project state
• daily reflection produced new durable learnings