Skip to content
Back to skills

Ccc Memory

ASecurity

Two-tier persistent memory for Claude Code sessions. Layer 1 is CLAUDE.md (always-loaded context). Layer 2 is a memory/ directory with dated notes, decisions, and…

  • 6 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added May 29, 2026
ai-agentsgosecurity

Works with

  • claude code

Security analysis

A100/100

Scanned May 29, 2026

npx -y skills add KevinZai/commander --skill ccc-memory --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ccc Memory?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ccc Memory
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/kevinzai-ccc-memory/badge)](https://www.skillsdirectory.com/skills/kevinzai-ccc-memory)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ccc-memory
description: "Two-tier persistent memory for Claude Code sessions. Layer 1 is CLAUDE.md (always-loaded context). Layer 2 is a memory/ directory with dated notes, decisions, and…"
model: sonnet
effort: medium
---

# /ccc-memory — Two-Tier Persistent Memory

Claude Code sessions forget everything at the end. Memory makes the important bits stick.

## When to use

- After resolving a tricky bug — save the root cause so next time the same class of issue is faster
- After making an architectural decision — record the reasoning so future-you doesn't re-argue with past-you
- When a project grows beyond one CLAUDE.md section — split into a memory/ tree
- Before a big /clear when you want to carry forward a specific insight

## Not for

- Session-scoped state that ends with this session — use /ccc-save-session for that
- Searchable cross-project knowledge — use /ccc-recall for that
- Per-skill notes — put those inline in the skill's body

## How it works

Two layers, different load patterns:

### Layer 1 — CLAUDE.md (always-loaded)

Every file Claude reads in a session is relative to CLAUDE.md. Put here:
- Project type + tech stack (once)
- Key paths + conventions
- Commands Kevin runs daily
- Cross-cutting style rules

Keep it tight — everything here costs context on every turn.

### Layer 2 — memory/ directory (on-demand)

Dated markdown files for things that don't belong in the always-loaded layer:
- `memory/YYYY-MM-DD-{slug}.md` — daily learnings, decisions, corrections
- `memory/decisions/` — one file per big call (why we picked Postgres over Mongo)
- `memory/gotchas/` — known footguns (tailwind v4 syntax, framer-motion v11 rename)
- `memory/preferences.md` — curated long-term prefs that didn't fit CLAUDE.md

Load on demand by reading the specific file — Claude sees only what's needed.

## Process

1. **Start of session:** `ls memory/` (if exists) → skim filenames → read the few that matter for today's task
2. **When you learn something:** add a dated entry to `memory/YYYY-MM-DD-{slug}.md`
3. **Corrections from Kevin:** append to `memory/preferences.md` with date — corrective framing ("You should X — are you doing it?")
4. **Weekly (or before big work):** consolidate — move frequently-referenced items from memory/ into CLAUDE.md; prune stale dated notes

## Example

After Kevin says "no more /ccc-style: prefix, use /ccc-*" — you write:

```
memory/2026-04-23-ccc-prefix-dash.md

## Correction: /ccc- not /ccc:

Date: 2026-04-23
Source: Kevin (chat)

All skill names use dash prefix /ccc-* not colon /ccc:*. Colon form is stale
from v3. Check: grep -rn "/ccc:" should return zero results outside CHANGELOG.

Test that guards this: commander/tests/dispatch-security.test.js — runKc('ccc:') rejected.
```

## Depends on

- `tasks/memory/` or `memory/` directory in project root (create if missing)
- Project CLAUDE.md as Layer 1

---

Adapted from `knowledge-work-plugins/productivity/memory-management` under Apache-2.0 license. Fitted to CC Commander's save-session / resume-session pattern.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.