compress-memory

Compress Memory

Purpose

Compress natural-language memory files into a terser form that preserves every byte a downstream consumer (skill or human) might key off. Reduces input tokens replayed every Claude Code session.

This is input token compression. Caveman-style output speech is explicitly NOT in scope (it would fight every skill in this suite that produces structured artifacts). See NOTICE.md for the relationship to caveman-compress.

When to Use

Explicit triggers

• /compress-memory <filepath> slash command
• Natural language: "compress STATE.md", "shrink CLAUDE.md", "compact this memory file"

Automatic trigger

• project-orchestration:pause-work invokes this skill on docs/planning/STATE.md (and ROADMAP.md if dirty) when docs/planning/ROADMAP.md frontmatter contains compress_memory: enabled.

When NOT to Use

This skill REFUSES the following files (see safety-rules.md for the full denylist):

• Any file under docs/plans/ — these are plan documents read literally by downstream skills
• *-design.md, *-impact-analysis*, *ui-contract*, *-review-*.md — contracts between skills
• MILESTONE.md — rewritten on milestone transitions, not pause-work
• *.original.md — backup files
• Anything not .md or .txt
• Files larger than 50 kB
• Files with compress: skip or compress_memory: skip in their frontmatter

When asked to compress a denied file, refuse with the specific reason. Do not modify the file.

Process

Step 1 — Validate the file

1. Resolve the file path. If a relative path was given, resolve against the project root.
2. Check the extension is .md or .txt. If not → abort with reason.
3. Check the path against the denylist in safety-rules.md. If denied → abort with the denylist pattern that matched and the user-facing message from safety-rules.md.
4. Read the file. If size > 50 kB → abort with reason.
5. If the file has YAML frontmatter and the frontmatter contains compress: skip or compress_memory: skip → abort with the per-file opt-out message.

Step 2 — Back up the original

1. Compute backup path: replace the file's .md (or .txt) extension with .original.md (or .original.txt). For docs/planning/STATE.md, the backup is docs/planning/STATE.original.md. NOT STATE.md.original.md.
2. If the backup file already exists → leave it untouched. The first backup is canonical.
3. If the backup file does NOT exist → write the original content to it before any modification of the source file.
4. If the backup write fails for any reason → abort WITHOUT modifying the original. Report the failure.

Step 3 — Generate the compressed candidate

Apply the rules in compression-rules.md:

1. Drop: articles, filler adverbs, pleasantries, hedging, connective fluff, imperative softeners.
2. Replace: verbose phrasing with shorter equivalents from the table.
3. Preserve byte-exact: fenced code blocks, indented code blocks, inline code spans, URLs and markdown links, file paths, shell commands, environment variables, version numbers, dates, frontmatter blocks, markdown tables (structure + cell text follows the same rules), headings (level + exact text), list nesting.

Apply rules to PROSE ONLY. Treat code blocks as read-only regions.

Step 4 — Validate the compressed candidate

Run all of the following mechanical checks BEFORE writing the candidate to disk. Each check compares the candidate against the original input.

1. Fenced code blocks — extract every fenced block from input and from candidate. Counts must match. Each block must be byte-equal.
2. Indented code blocks — extract every 4-space indented code block. Counts must match. Each block must be byte-equal.
3. Inline code spans — count must match.
4. URLs — extract the full set of URLs from input and from candidate. Sets must be equal (no additions, no removals).
5. Markdown headings — extract every heading line. Count, exact text, and order must match.
6. Frontmatter — if input has a frontmatter block, candidate must have a byte-equal frontmatter block. If input has no frontmatter, candidate must also have none.
7. Markdown tables — count of table blocks must match; column count per table must match.

If ANY check fails:

• Do NOT write the candidate to disk.
• Confirm the original file still matches the backup (it should — we haven't written anything).
• Report the specific failure to the user.
• If invoked from pause-work, return failure to the parent so it can log and continue with the uncompressed file.

Step 5 — Write the compressed file

If all validation checks pass:

1. Write the compressed candidate over the original file path using the Write tool.

2. Verify the write succeeded by re-reading the file and confirming size matches the candidate.

3. Report to the user:

   Compressed <path>:

   • Before: <X> bytes
   • After: <Y> bytes
   • Saved: <Z>%
   • Backup: <basename>.original.md (e.g. docs/planning/STATE.md → docs/planning/STATE.original.md)

Decision Tree

digraph compress_memory {
  "Request to compress <file>" [shape=doublecircle];
  "Allowed extension?" [shape=diamond];
  "Path on denylist?" [shape=diamond];
  "Size <= 50 kB?" [shape=diamond];
  "Frontmatter opts out?" [shape=diamond];
  "Backup exists?" [shape=diamond];
  "Write backup" [shape=box];
  "Generate compressed candidate" [shape=box];
  "Validation passes?" [shape=diamond];
  "Write compressed file" [shape=box];
  "Report result" [shape=doublecircle];
  "Abort with reason" [shape=doublecircle];

  "Request to compress <file>" -> "Allowed extension?";
  "Allowed extension?" -> "Abort with reason" [label="no"];
  "Allowed extension?" -> "Path on denylist?" [label="yes"];
  "Path on denylist?" -> "Abort with reason" [label="yes"];
  "Path on denylist?" -> "Size <= 50 kB?" [label="no"];
  "Size <= 50 kB?" -> "Abort with reason" [label="no"];
  "Size <= 50 kB?" -> "Frontmatter opts out?" [label="yes"];
  "Frontmatter opts out?" -> "Abort with reason" [label="yes"];
  "Frontmatter opts out?" -> "Backup exists?" [label="no"];
  "Backup exists?" -> "Generate compressed candidate" [label="yes"];
  "Backup exists?" -> "Write backup" [label="no"];
  "Write backup" -> "Generate compressed candidate";
  "Generate compressed candidate" -> "Validation passes?";
  "Validation passes?" -> "Abort with reason" [label="no"];
  "Validation passes?" -> "Write compressed file" [label="yes"];
  "Write compressed file" -> "Report result";
}

Common Mistakes

1. Modifying code inside fenced blocks — even "obvious cleanups" inside triple-backtick fences are forbidden. Treat code blocks as read-only.
2. Compressing tables by deleting columns — table structure (pipe layout, alignment row, column count) is preserved. Only cell prose follows drop/replace rules.
3. "Just one URL changed" — that's a failure — URL set equality is strict. A typo'd URL is a validation failure, not a successful compression.
4. Re-overwriting *.original.md — the first backup is canonical. Do not re-create the backup on subsequent compressions; the user wants to be able to recover the pristine version.
5. Operating on a denylisted file because the user asked — denylist is hard-coded for a reason. Refuse, explain why, point at the safer files.
6. Appending .original.md to the full filename — backup naming REPLACES the extension. STATE.md → STATE.original.md, NOT STATE.md.original.md. Getting this wrong silently violates the denylist pattern matching.

References

• compression-rules.md — the drop/replace/preserve rules in detail
• safety-rules.md — denylist, backup invariant, validation
• NOTICE.md — attribution to caveman (MIT)