memex: Skill for Claude Code

Session Summarization Workflow

For NEW sessions:

1. Read the file — parse each line as JSON
2. Filter to user/assistant messages — skip file-history-snapshot, tool results, subagent internals
3. Strip noise — remove IDE tags (<ide_opened_file>), command outputs (<local-command-stdout>), system reminders, task notifications
4. Identify the project — from cwd, project field, or folder encoding
5. Understand the task — what was the user trying to accomplish?
6. Document outcomes — what was built, changed, fixed, or decided?
7. Note technologies — languages, frameworks, tools used
8. Write summary — follow @SESSION_TEMPLATE.md
9. Update manifest — add entry to projects/.manifest.json

For UPDATED sessions (session continued since last summary):

1. Read the manifest — find lines count from last processing
2. Read the JSONL file — focus on lines AFTER the previously processed count
3. Read the existing summary — understand what's already documented
4. Identify new work — what happened in the continuation?
5. Update the summary — append new sections or extend existing ones. Do NOT rewrite what's already there unless it's wrong.
6. Update manifest — set lines to current count

Output Format

Follow the template in @SESSION_TEMPLATE.md

Manifest Update

After processing, update projects/.manifest.json:

manifest[rel_path] = {
    "lines": current_line_count,
    "last_timestamp": last_message_timestamp,
    "summary": "projects/ProjectName/YYYY-MM-DD_description.md",
    "status": "summarized"  # or "skipped" for trivial sessions
}

Rules

• Be factual, not verbose. Summarize what happened, not what Claude said.
• Include enough detail that someone reading the summary months later can understand the work.
• If the session involves Chinese text, preserve key Chinese terms where they add context.
• If the session is trivial (<5 real messages, no meaningful work), mark as "skipped" in manifest.
• Always update the project README.md session table and projects/INDEX.md.