harvest-memory

Harvest Memory

Read CLAUDE.md with full comprehension, classify sections intelligently, and store high-value entries in the memory system using mcp__memory__* tools.

Usage

• /harvest-memory — harvest CLAUDE.md in current working directory
• /harvest-memory --all — harvest CLAUDE.md from every project under ~/Code/

Steps

1. Locate CLAUDE.md: Check ./CLAUDE.md, then ./.claude/CLAUDE.md. If --all flag is provided, glob ~/Code/*/CLAUDE.md and process each project in turn.

2. Derive scope: Use project:<basename-of-cwd> as the scope (e.g., project:kranz-tv).

3. Parse sections: Split the file by ## headings. For each section, classify as high-value or skip:

   Section heading contains	Entity suffix	entity_type	Skip?
   Architecture	architecture	decision	No
   Scheduling	scheduling	decision	No
   Data flow	data-flow	decision	No
   Key hooks	key-hooks	decision	No
   Routing	routing	decision	No
   Channel import	channel-import	decision	No
   Constraints	constraints	decision	No
   Key Patterns	key-patterns	decision	No
   Tech Stack	tech-stack	project	No
   Observability	observability	service	No
   Commands	—	—	Yes
   Environment Variables	—	—	Yes
   Testing	—	—	Yes
   TypeScript	—	—	Yes

4. Check existing entries: For each high-value section, call mcp__memory__memory_get_entity with name=<project>-<suffix> and scope=project:<project>. Compare stored content to current section content.

5. Store or update:

   • If no existing entry → call mcp__memory__memory_store with confidence=0.95 and source='skill:harvest-memory'
   • If content has changed → call mcp__memory__memory_store to update (the tool upserts)
   • If content is identical → skip (note as "unchanged" in output)

6. Output results table:

   | Entity Name              | Type     | Action           | Preview (60 chars)                  |
   |--------------------------|----------|------------------|-------------------------------------|
   | kranz-tv-architecture    | decision | created          | KranzTV is a retro cable TV experi… |
   | kranz-tv-scheduling      | decision | updated          | getSchedulePosition(channel, times… |
   | kranz-tv-tech-stack      | project  | unchanged        | —                                   |
   

Key Differences from Hook

The SessionEnd hook harvests sections via raw awk text extraction into SQLite directly. This skill uses Claude's full comprehension to:

• Understand context and nuance in each section
• Write cleaner, more useful summaries if the raw text is too long
• Cross-reference sections for relationships
• Detect implicit constraints not marked with a specific heading

Use source tag skill:harvest-memory so hook entries (tagged hook:harvest) coexist without conflict.

Notes

• Trim section content to 1000 chars max before storing
• Ignore frontmatter and top-level # heading
• If CLAUDE.md is missing entirely, report clearly and exit
• For --all mode, process each project directory and report a combined results table