REM Sleep - Memory Extraction

Purpose

Scan the current conversation for valuable information worth persisting, categorize it, and dispatch the memory-extractor agent to write it to the memories table.

Trigger

• Manual: user invokes /rem-sleep
• Automatic: PreCompact hook outputs a reminder, Claude invokes this skill

Data Storage Rules

• NEVER write memory data to local files — all memory goes to the memories table
• Use ToolSearch to discover Supabase MCP tools for INSERT/UPDATE operations
• Temp files in /tmp/ are acceptable for transient processing

Execution

1. Determine Project Name

git remote get-url origin 2>/dev/null | sed 's|.*/||;s|\.git$||' || basename "$PWD"

Normalize to lowercase with hyphens.

2. Scan Conversation

Review the entire conversation for extractable items in these categories:

Category	What to Look For	DB category	DB subcategory
Decisions	Architectural choices, library selections, pattern choices	decision or project	NULL or <project>
Preferences	User's stated preferences for style, tooling, workflow	preference	NULL or topic
Patterns	Recurring approaches, idioms, code patterns used	knowledge	<topic>
Issues	Bugs found, workarounds discovered, gotchas	project	<project>
Learnings	New knowledge gained, framework behaviors, API quirks	knowledge	<topic>

3. Categorize and Deduplicate

For each extracted item:

• Assign category and subcategory
• Write a concise key (slugified, 2-6 words, e.g., zustand-over-redux)
• Distill content (value field) to factual essence (no commentary)
• Assign confidence level:
  • high: explicitly stated by user or confirmed through implementation
  • medium: inferred from conversation patterns
  • low: mentioned once, possibly context-dependent
• Assign relevant tags
• Set source to rem-sleep

4. Dispatch Memory Extractor

Launch the memory-extractor agent with the categorized items:

Use the Task tool with subagent_type "general-purpose" to run the memory-extractor agent.
Provide the list of categorized memory items as structured input.

5. Report Results

If manually invoked (/rem-sleep): Output a brief summary:

Memories extracted:
- 2 decisions (project-specific)
- 1 preference (cross-project)
- 1 knowledge item (TypeScript patterns)

If hook-triggered (PreCompact): Operate silently. Do not output a summary unless errors occur.

What NOT to Extract

• Transient debugging steps (unless they reveal a persistent gotcha)
• Code that was written and immediately discarded
• Questions the user asked that were fully resolved in-session
• Implementation details that are visible in the code itself
• Anything already stored in the memories table (check first)

Examples

Good Extractions

• "User prefers Zustand over Redux for state management" → category=preference, key=zustand-over-redux
• "This project uses tRPC for API layer with Zod validation" → category=project, subcategory=my-app, key=trpc-api-layer
• "Always run migrations before seeding in this project" → category=project, subcategory=my-app, key=migrations-before-seeding
• "React 19 useActionState replaces useFormState" → category=knowledge, subcategory=react, key=useactionstate-replaces-useformstate

Bad Extractions

• "Fixed the typo on line 42" → too transient
• "The button component renders a <button> element" → obvious from code
• "User asked how to use useState" → resolved in session