Memory Distill

Memory Distill

Utility skill. Captures user preferences and writing feedback into persistent memory.

Memory Architecture

• L1 — ~/.autocrew/MEMORY.md (≤120 lines): Working memory. Brand profile, audience, account status, writing preferences summary.
• L2 — ~/.autocrew/memory/: Detailed logs and digests. Referenced on demand.

When to Activate

• User approves or rejects a topic
• User edits a draft significantly
• User gives explicit feedback ("太正式了", "多用emoji", "短一点")
• User shares performance data ("这条笔记1000赞")

Steps

Feedback Capture

1. Identify the feedback signal:

   • Approval: user says "好" / "可以" / "就这个" → positive signal
   • Rejection: user says "不行" / "太..." / "换一个" → negative signal
   • Edit: user provides a rewrite → compare original vs edit
   • Performance: user shares metrics → high/low performance signal

2. Analyze what the user liked or disliked. Be specific:

   • Bad: "user likes emoji"
   • Good: "user prefers 3-5 emoji per XHS post, placed at paragraph starts, favorites: 🔥 💡 ✨"

3. Formulate a learning as a concise preference statement.

4. Read ~/.autocrew/MEMORY.md. Append the learning to the appropriate section:

Prefer using autocrew_memory:

{
  "action": "capture_feedback",
  "content_id": "content-xxx",
  "signal_type": "edit",
  "feedback": "太正式了",
  "modified_text": "用户改后的版本"
}

## Writing Preferences
- [Date] XHS posts: user prefers casual tone, 3-5 emoji, short paragraphs
- [Date] Titles: user likes curiosity-gap style, dislikes clickbait

## Content Edit Preferences
- [Date] Shortened opening from 3 sentences to 1 — user prefers immediate hooks
- [Date] Replaced formal "综上所述" with casual "所以说" — user wants conversational tone

## Performance Insights
- [Date] "AI工具真香清单" got 1000 likes on XHS — list format + "真香" hook works

Capacity Check

If MEMORY.md exceeds 150 lines:

1. Identify entries that can be consolidated (similar learnings → one summary)
2. Move detailed case studies to ~/.autocrew/memory/archive-{date}.md
3. Replace with a one-line summary in MEMORY.md

Guidelines

• Be specific — actionable preferences, not vague observations
• Don't overwrite previous learnings, accumulate them
• Date every entry for tracking evolution
• Consolidate when file gets long, don't let it grow unbounded

Error Handling

Changelog

• 2026-03-31: v1 — Adapted from Qingmo memory-distill.md v3. Simplified to two-layer architecture (no LanceDB dependency). File-based storage only.