harrylettering/codememory/codememory-summarization-skill

CodeMemory Summarization Skill

Guidelines for creating high-quality conversation summaries in CodeMemory's DAG structure.

When This Skill Applies

This skill activates when:

• Creating or reviewing conversation summaries
• Working with the DAG compaction system
• Discussing summary quality or token efficiency
• Optimizing the summary hierarchy
• Training or configuring the summarization agent

Summary Quality Guidelines

What to Include in Every Summary

1. Key Decisions: Any decisions made, with rationale
2. Action Items: Tasks assigned, with owners and deadlines
3. Technical Details: Architecture choices, API designs, schema changes
4. Problem Statements: What issues were being addressed
5. Alternatives Considered: Options that were rejected and why
6. Ambiguities Resolved: Questions that were clarified
7. User Preferences: Stated preferences that affect future work

What to Exclude

• Redundant chit-chat
• Failed tool calls (unless the failure is informative)
• Repetitive restatements
• Simple acknowledgments ("ok", "got it")

Leaf Summary Guidelines

Target: 8-16 messages, ~1200 tokens

Format:

## Summary
 
[Concise overview of the conversation segment]
 
### Key Decisions
-[Decision 1]: [What was decided, why]
-[Decision 2]: [What was decided, why]
 
### Action Items
-[ ] [Action 1]: [Owner/Deadline if mentioned]
-[ ] [Action 2]: [Owner/Deadline if mentioned]
 
### Technical Details
[Relevant technical information, code snippets, architecture]

Condensed Summary Guidelines

Target: 2-4 child summaries, ~2000 tokens

Format:

## Summary
 
[Higher-level synthesis of the child summaries, showing how they connect]
 
### Key Themes
-[Theme 1]: [How it appears across child summaries]
-[Theme 2]: [How it appears across child summaries]
 
### Major Outcomes
-[Outcome 1]: [Result from this phase]
-[Outcome 2]: [Result from this phase]
 
### Child Summaries Referenced
-sum_abc123: [Brief description]
-sum_def456: [Brief description]

DAG Structure Best Practices

1. Balanced Tree: Aim for 2-4 children per condensed node
2. Logical Boundaries: Split conversations at natural stopping points
3. Topic Cohesion: Each summary should have a clear, single topic
4. Forward References: If a later summary references an earlier one, note it
5. Token Awareness: Stay within target token counts

Quality Checklist

Before marking a summary complete:

• [ ] All key decisions are included
• [ ] All action items are captured
• [ ] Technical details are accurate
• [ ] Summary is concise but comprehensive
• [ ] Token count is within target range
• [ ] No important information is lost
• [ ] Parent/child relationships make sense