Persistent Memory — AI skill for Claude Code

persistent-memory

**One reviewed local memory layer for compatible AI agents.**

Claude does not automatically know what you told Codex, and Codex does not automatically know what you saved elsewhere. Persistent Memory gives compatible agents a shared, readable source of context instead of asking you to paste it again.

It complements platform memory; it does not replace it, import your old chats, or run in the background.

What It Does

• Loads baseline context — load memory reads your core files and index, not every project file.
• Saves with review — the AI proposes the content and destination; you approve before it writes.
• Coordinates wrap-up updates — one request lets the agent decide which project-context layers actually need changes.
• Loads details on demand — relevant project or note files are read only when the topic needs them.
• Uses optional Summary-first loading — for an active file with a reviewed ## Summary, the agent reads the short summary first and opens the full file only when the request needs more detail. Files without a summary keep the existing behavior.
• Manages lifecycle safely — archive, trash, recovery, and memory health are confirmation-based.
• Stays transparent — memory is plain local Markdown you can inspect and edit.

Quick Start

1. Install the skill for an agent that can discover Skills.
2. Start with an explicit command: remember this or 记住这个.
3. Review the AI's proposed file and wording, then confirm it.
4. In a later compatible conversation, say load memory or 加载记忆.

If you already have a memory folder created before lifecycle support, run `memory upgrade` / `升级记忆` before using archive, delete, recover, or memory health.

How It Works

~/.persistent-memory/
├── _core/                 # Baseline identity and collaboration context
├── _index.md              # One-line routes to active on-demand memory
├── projects/              # Active project context
├── notes/