Summarization

Purpose / When to Activate

Activate when:

• Session memory grows large
• Decisions accumulate across sessions
• Project context becomes fragmented
• Memory retrieval returns too many files
• Token usage increases noticeably

This skill is both preventive and corrective.

Process

Step 0: Enumerate categories. Read contexts/memory/index.md. List all registered categories. For each category, assess whether summarization is warranted: trigger if >5 files in the category OR estimated token count >5000 tokens OR memory-retrieve is returning too many results for this category. Skip categories that do not meet any threshold.

For each category that meets the threshold:

1. Identify durable information — extract confirmed decisions, stable constraints, validated assumptions, current priorities, key outcomes, known risks. Ignore brainstorming noise, intermediate reasoning, abandoned ideas.

2. Compress into structured summary — use the template below. Prefer bullets over prose.

3. Archive raw memory — move original files to contexts/memory/archive/{category}-{YYYY-MM-DD}.archive.md. Concatenate multiple files from the same category into a single archive file. Only summaries remain active.

4. Update memory index — record new summary files, archived sources, affected categories.

Summary File Template

# {Category} — Summary
> Summarized from {N} files on {YYYY-MM-DD}
> Compression: ~{original_tokens} → ~{summary_tokens} ({percentage}%)

## Decisions
- {bullet per confirmed decision}

## Constraints
- {bullet per active constraint}

## Priorities
- {bullet per current priority}

## Open Questions
- {bullet per unresolved question, if still relevant}

Outputs

• contexts/memory/summaries/{category}.summary.md — compressed, structured summary for each processed category
• contexts/memory/archive/{category}-{YYYY-MM-DD}.archive.md — concatenated originals, retained for audit
• contexts/memory/index.md — updated to reflect new summaries and archived sources

Quality Checks

A good summary:

• Summary token count is ≤20% of original — verify before replacing active memory
• Every summarized category has its archive file created before the summary replaces it
• Preserves all actionable knowledge
• Removes all conversational fluff
• Supports future decisions without rereading history

Non-Goals

This skill must NOT:

• Invent new knowledge
• Reinterpret decisions
• Remove active constraints
• Keep long narrative text

Distill knowledge. Delete noise. Small, sharp context always beats full history.