Agent Skills: Memory Defragmentation

Memory Defragmentation

Requires context-repo extension with memory at .pi/memory/

Splits large, multi-purpose memory blocks into focused single-purpose files with hierarchical / naming.

When to Use

• Memory blocks have redundant information
• Files mix multiple unrelated topics
• Memory lacks structure (walls of text)
• After major project milestones
• Every 50-100 conversation turns

Workflow

Step 1: Backup (MANDATORY)

Use the memory_backup tool before proceeding. This is your safety net.

Step 2: Analyze Current Memory

find .pi/memory -name '*.md' | while read f; do
  echo "=== $f ($(wc -l < "$f") lines) ==="
  head -5 "$f"
  echo
done

For each file, determine:

• Does it serve 2+ distinct purposes? → needs splitting
• Is it >40 lines? → candidate for splitting
• Does it overlap with another file? → consolidate

Step 3: Decompose

Split multi-purpose blocks into focused files using hierarchical naming:

Before:

system/project.md  (80 lines mixing overview, tooling, conventions, gotchas)

After:

system/project/overview.md
system/project/tooling.md
system/project/conventions.md
system/project/gotchas.md

Use memory_write for each new file, then delete the original:

rm .pi/memory/system/project.md

Step 4: Clean Up

For each file (new and existing):

• Add markdown structure (headers, bullets)
• Remove redundancy across files
• Remove speculation ("probably", "maybe")
• Keep only actionable, concrete information
• Resolve contradictions

Step 5: Commit

memory_commit({ message: "refactor: defragment memory — split N files into M focused blocks" })

Step 6: Report

Provide a summary:

• Files created (new decomposed blocks)
• Files modified (what changed)
• Files deleted (if any, explain why)
• Before/after file counts and line counts

Evaluation Criteria

1. DECOMPOSITION — Each file has ONE clear purpose described by its filename
2. STRUCTURE — Headers, bullets, scannable at a glance
3. CONCISENESS — No redundancy, no speculation, only unique value
4. CLARITY — Contradictions resolved, plain language, actionable
5. ORGANIZATION — General to specific within files, important first

Naming Rules

• Use / hierarchy: project/tooling/testing.md (not project-tooling-testing.md)
• 2-3 levels of nesting
• ~40 lines max per file
• Descriptive frontmatter descriptions

What to Preserve

• User preferences (sacred — never delete)
• Project conventions discovered through experience
• Important context for future sessions
• Learnings from past mistakes

Rollback

If something goes wrong:

/memory-backups     # list available backups
/memory-restore <backup-name>