claude-memory-sync

claude-memory-sync

A Claude Code plugin that keeps your project's memory files (CLAUDE.md, .claude/rules/*.md, and related files) synchronized with the actual state of your codebase.

---

The problem

Claude's memory files go stale. You add a new environment variable, rename a module, remove a deprecated endpoint - but CLAUDE.md still describes the old world. Over time, Claude works from an increasingly inaccurate picture of your project, missing context it should have and confidently referencing things that no longer exist.

What this skill does

When you ask Claude to sync its memory, this skill:

1. Finds your memory files - CLAUDE.md at any depth, .claude/rules/*.md, and any other files they cross-reference.
2. Establishes a baseline - reads each file's last-sync timestamp to know exactly how far back to look for changes.
3. Detects what changed - runs both a git log sweep and a filesystem mtime scan, then merges and deduplicates the results.
4. Reads the diffs - classifies each change as trivial or memory-relevant and summarizes the impact.
5. Maps changes to memory files - figures out which sections of which files are affected, respecting scope frontmatter that restricts rules files to specific parts of the codebase.
6. Verifies each file - goes section by section checking accuracy, staleness, and gaps.
7. Proposes a change list - additions, updates, removals, and moves.
8. Safety-gates large removals - presents risky changes as a single consolidated list for your review before writing anything.
9. Applies edits surgically - minimum changes, preserving your existing style and structure.
10. Creates new context files - if a new subsystem is large enough to warrant its own .claude/rules/<domain>.md, it creates one.
11. Runs a cross-file consistency review - checks for contradictions, orphaned references, duplicated content, and scope mismatches.
12. Reports what it did - a concise summary of every change made, removed, and flagged.

The core principle: Claude's memory files must reflect the project as it IS, not as it WAS.

---

Installation

From GitHub (recommended)

/plugin marketplace add <your-username>/claude-memory-sync
/plugin install claude-memory-sync@claude-memory-sync

Then reload:

/reload-plugins

Manual install (no GitHub required)

Copy the skill folder directly into your global skills directory:

cp -r plugins/claude-memory-sync/skills/claude-memory-sync ~/.claude/skills/

---

Usage

Once installed, trigger the skill with natural language in any Claude Code session:

sync Claude memory
update CLAUDE.md
refresh Claude's project knowledge
check if Claude files are up to date
the project has changed, Claude doesn't know about it
run a memory audit

Claude will run all phases automatically and present a sync report at the end.

Force a full review

If you want Claude to re-verify everything even if nothing appears to have changed:

force a full Claude memory sync

---

How the timestamp system works

Every memory file managed by this skill ends with a metadata block:

---
<!-- claude-memory-sync: 2026-03-18T14:00:00Z -->

This timestamp records when the file was last verified. On the next sync, the skill uses the oldest timestamp across all your memory files as the baseline for change detection - so if any file is behind, the entire change window expands to cover it.

Files without a timestamp are treated as never-synced and receive a full re-verification pass regardless of what else changed in the project.

After every sync, the timestamp is updated on every file that was touched.

---

What gets updated

The skill handles all standard Claude Code memory file locations:

File / pattern	Description
CLAUDE.md	Root and subdirectory project context files
.claude/rules/*.md	Scoped rules files with optional frontmatter
.claude/*.md	Other top-level Claude context files
Referenced files	Any .md paths mentioned in the above files

Scope-aware rules files

Rules files can declare their scope using YAML frontmatter:

---
applies_to: ["src/frontend/**", "*.tsx"]
context: frontend
---

# Frontend Rules
...

The skill respects this - if only backend files changed, a frontend-scoped rules file is left alone. Files without frontmatter are treated as global and always included.

---

Safety model

The skill distinguishes between changes that are safe to apply automatically and those that need your review.

Applied without confirmation:

• Adding a new env var to the documented list
• Adding a new script to the commands section
• Documenting a new top-level directory
• Any clearly additive, unambiguous change