Memory

Memory - Persistent Knowledge System

Purpose: Maintain persistent knowledge about projects, environments, and user preferences so you never have to repeat yourself.

Core Principle

If you tell me something once, I should remember it forever.

This skill captures and organizes:

• Project-specific facts (hosting, domains, tech stack, integrations)
• Environment details (API keys locations, deployment workflows)
• User preferences (coding style, communication preferences)
• Decisions and their rationale
• Lessons learned from debugging sessions

Memory Structure

All memories stored in ${PAI_DIR}/memory/:

memory/
├── projects/           # Project-specific knowledge
│   ├── carbene-website.md
│   ├── carbene-homelab.md
│   ├── pai.md
│   └── cyberdefense-tactics.md
├── user/               # User preferences and personal info
│   └── preferences.md
└── global/             # Cross-project knowledge
    └── integrations.md

Workflows

1. Capture Learning (Proactive)

When to use: Automatically after learning something new about a project or user.

TRIGGER: Any of these situations:
- User corrects a mistake ("It's Vercel, not Cloudflare")
- User provides new information ("The domain is carbene.ai")
- Discovering project structure or integrations
- User states a preference
- Resolving a confusing issue

ACTION:
1. Identify the memory category (project/user/global)
2. Read existing memory file (if exists)
3. Append new fact with timestamp
4. Confirm capture to user (briefly)

2. Recall Context (On Session Start)

When to use: At the beginning of each session, based on working directory.

ACTION:
1. Detect current working directory
2. Map to project memory file
3. Load relevant memories into context
4. Silently apply (no need to announce)

3. Query Memory

When to use: When user asks "What do you know about X?"

ACTION:
1. Search memory files for relevant content
2. Return organized summary
3. Offer to update if anything is outdated

Memory Format

Each memory file uses this format:

# Project: [Name]

## Quick Facts
| Key | Value |
|-----|-------|
| Domain | example.com |
| Hosting | Vercel |
| Repo | github.com/user/repo |

## Tech Stack
- Framework: Next.js 16
- Package Manager: Bun
- Styling: Tailwind CSS

## Integrations
- **Google Sheets**: Contact form and assessment data
- **Apps Script**: Email automation at `docs/google-apps-script.js`

## Key Files
- `CLAUDE.md` - Project instructions
- `.env.local` - Environment variables
- `docs/google-apps-script.js` - Email automation code

## Decisions & Rationale
- [2026-01-20] Using Vercel instead of Cloudflare Pages for easier deployment
- [2026-01-20] Assessment tracks by index not value due to duplicate scores

## Lessons Learned
- [2026-01-20] Apps Script requires NEW deployment after changes (saving isn't enough)
- [2026-01-20] carbene.ai is production, yourdomain.com is homelab/internal - never confuse

## Notes
Additional context that doesn't fit elsewhere.

Directory-to-Project Mapping

| Directory Pattern | Project Memory | |-------------------|----------------| | */carbene-ai-website* | projects/carbene-website.md | | */homelab* | projects/carbene-homelab.md | | */PAI* or */.claude* | projects/pai.md | | */cyberdefense* | projects/cyberdefense-tactics.md |

Auto-Capture Triggers

The Memory skill should automatically capture when:

1. Correction detected: User says "no", "actually", "it's not X, it's Y"
2. New fact provided: User shares hosting, domain, credentials location
3. Integration discovered: Finding how systems connect
4. Decision made: Choosing one approach over another
5. Bug resolved: Understanding why something failed

Commands

• /memory show [project] - Display memories for a project
• /memory add [project] [fact] - Manually add a memory
• /memory search [query] - Search across all memories
• /memory projects - List all tracked projects

Integration with Session Hooks

The SessionStart hook should:

1. Detect working directory
2. Load matching project memory
3. Apply context silently

The SessionEnd hook should:

1. Review session for new learnings
2. Prompt to capture important facts
3. Update relevant memory files

Example Memories to Capture

From today's session:

• CarbeneAI website hosted on Vercel (not Cloudflare)
• Production domain: carbene.ai
• Homelab domain: yourdomain.com
• Google Apps Script in docs/google-apps-script.js
• Notification email: you@yourdomain.com
• Assessment has 10 questions with duplicate score values
• Apps Script requires new deployment after code changes

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: CarbeneAI
• Source: CarbeneAI/Forge
• License: MIT
• Homepage: http://carbene.ai/open-source

Install and usage instructions live in the source repository linked above.