How do I install the Memory Audit skill?

Run npx skillmds@latest add adamradek-ux/memory-audit in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Memory Audit safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Memory Audit?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Memory Audit free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Memory Audit?

adamradek-ux (@adamradek-ux) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Memory Audit by adamradek-ux — Agent Skill

Name: Memory Audit
Author: adamradek-ux

Memory Audit

Compares two memory layers — git-backed context notes in your vault vs the non-git Claude Code auto-memory store — finds overlap/drift and convention drift, and PROPOSES a source of truth plus what to merge or delete. Read-only analysis; propose → OK → write. Triggers: "memory audit", "what's duplicated in memory", "sync my memory", "what's stale in my context notes", and Czech aliases "audit paměti", "co je zastaralé v paměti".

adamradek-ux f9ce346 2.8 KB Updated Sep 17, 2026

File contents

Check memory consistency across two layers and propose fixes. The analysis is read-only; any write/delete goes through propose → OK → write.

The two memory layers (context)

Vault context notes (e.g. context/ or _context/ in your notes repo) — git-backed, human-facing (who you are, current focus, decisions log, lessons). Travels with your git backup.
Claude Code auto-memory store — ~/.claude/projects/<project-slug>/memory/ (auto-loaded via MEMORY.md; file prefixes like feedback_ / reference_ / project_ / user_). Not covered by your git backup → risk of loss.

Procedure

1. Load both layers

Vault: your context/notes files, lessons, decisions log.
Memory store: the MEMORY.md index + the individual *.md files.

2. Find 4 classes of problems

Class	What to look for
Overlap/drift	Same topic in both layers with diverging content. Which one is the source of truth?
Staleness	A `feedback_`/`reference_` claim that no longer holds (verify against the code/docs/reality — memory is point-in-time, not live).
Index orphan	A `memory/*.md` file without a line in `MEMORY.md` (it never loads), or a `MEMORY.md` line without a file.
Convention drift	Two competing conventions for the same thing (e.g. old vs new daily-note location). Which is current?

3. Verify before proposing

Verify every "stale" claim against reality (code/file/docs), not against a hunch. What cannot be verified, mark "to confirm with the user" — do not delete it.
For overlaps, pick the source of truth by the layer's role (durable technical fact → memory store; human-facing state/decisions → vault). Don't guess — show evidence.

4. Propose the fix (propose → OK → write)

For each finding: file + problem + concrete action (merge into X, delete duplicate, add the missing MEMORY.md pointer, correct the stale claim, decide the convention). Order by risk (data loss > stale fact > cosmetics). Wait for the OK.

Deleting memory is irreversible and not covered by git — for every delete proposal, ask for an explicit OK separately.

What this skill does NOT do

It does not delete or rewrite anything without an OK.
It does not lint your vault structure (use your own linter or hooks for that) — it focuses on memory consistency across the two layers.

---name: memory-auditdescription: Compares two memory layers — git-backed context notes in your vault vs the non-git Claude Code auto-memory store — finds overlap/drift and convention drift, and PROPOSES a source of truth plus what to merge or delete. Read-only analysis; propose → OK → write. Triggers: "memory audit", "what's duplicated in memory", "sync my memory", "what's stale in my context notes", and Czech aliases "audit paměti", "co je zastaralé v paměti".---Check memory consistency across two layers and **propose** fixes.The analysis is read-only; any write/delete goes through propose → OK → write.## The two memory layers (context)- **Vault context notes** (e.g. `context/` or `_context/` in your notes repo) — git-backed, human-facing (who you are, current focus, decisions log, lessons). Travels with your git backup.- **Claude Code auto-memory store** — `~/.claude/projects/<project-slug>/memory/` (auto-loaded via `MEMORY.md`; file prefixes like `feedback_` / `reference_` / `project_` / `user_`). **Not covered by your git backup** → risk of loss.## Procedure### 1. Load both layers- Vault: your context/notes files, lessons, decisions log.- Memory store: the `MEMORY.md` index + the individual `*.md` files.### 2. Find 4 classes of problems| Class | What to look for ||---|---|| **Overlap/drift** | Same topic in both layers with diverging content. Which one is the source of truth? || **Staleness** | A `feedback_`/`reference_` claim that no longer holds (verify against the code/docs/reality — memory is point-in-time, not live). || **Index orphan** | A `memory/*.md` file without a line in `MEMORY.md` (it never loads), or a `MEMORY.md` line without a file. || **Convention drift** | Two competing conventions for the same thing (e.g. old vs new daily-note location). Which is current? |### 3. Verify before proposing- Verify every "stale" claim against reality (code/file/docs), not against a hunch. What cannot be verified, mark "to confirm with the user" — do not delete it.- For overlaps, pick the source of truth by the layer's role (durable technical fact → memory store; human-facing state/decisions → vault). Don't guess — show evidence.### 4. Propose the fix (propose → OK → write)For each finding: **file + problem + concrete action** (merge into X, deleteduplicate, add the missing `MEMORY.md` pointer, correct the stale claim, decide theconvention). Order by risk (data loss > stale fact > cosmetics). Wait for the OK.**Deleting memory is irreversible and not covered by git** — for every deleteproposal, ask for an explicit OK separately.## What this skill does NOT do- It does not delete or rewrite anything without an OK.- It does not lint your vault structure (use your own linter or hooks for that) — it focuses on memory consistency across the two layers.

Frequently asked questions