One memory for all your AI agents.
Claude doesn't know what you told Cursor. Codex doesn't know what Antigravity saved. Switch tools and you start from zero.
Install this skill and every agent reads the same memory. One command. No copy-pasting. No repeating yourself.
You: load memory
AI: Context loaded. I know you're a product designer at a startup, currently focused on the v2 redesign. You prefer direct feedback and like to think through problems step by step before coding. I also have notes on 3 projects and your design principles.
You: Let's pick up the onboarding flow we discussed last week.
AI: [reads the relevant project file on demand] Right — last time we narrowed it down to two approaches...
That's the experience. One command, and your AI picks up where you left off — no matter which agent you're using.
Every agent has some form of memory now. The problem isn't that they can't remember — it's that they each remember separately. Your Claude doesn't know what you told Cursor. Your Codex doesn't know what Antigravity saved. Switch tools and you start from zero.
This skill replaces all of that with one shared memory on your local machine. You control what's saved, you can read and edit every file, and it works across every agent that supports the skills protocol.
- Install the skill
- Start a conversation and share something about yourself
- The AI will ask: "Want me to save this to your memory?"
- Say yes — that's it. Your memory system is set up.
Next time, just say "load memory" and your AI knows you.
- Load your context instantly — say "load memory" and your AI picks up right where you left off
- Save what matters — the AI suggests saving important info, you decide what to keep
- Two-tier architecture — essential context (who you are, how you work) loads every time; project details load only when relevant
- You stay in control — nothing is saved without your review and approval
~/.persistent-memory/ ├── _core/ # Always loaded (identity, preferences) ├── _index.md # Lightweight index of all other files ├── projects/ # Loaded on demand └── notes/ # Loaded on demand
Core files load every conversation — your AI always knows the basics. Everything else is indexed with one-line summaries and loaded only when the topic comes up. This keeps token usage low as your memory grows.
| Say this | What happens |
|---|---|
| "load memory" | Loads your full context |
| "remember this" / "save this" | Saves info to memory (with your review) |
| "update memory" | AI scans the conversation and suggests what to save |
| "memory status" | Shows all saved files and their summaries |
The default structure (projects/, notes/) is just a starting point. You can:
- Rename directories to match your workflow (e.g.,
clients/,courses/) - Add new directories anytime — they'll be indexed automatically
- Reorganize
_core/files to fit what you need in every conversation
The only fixed parts are _core/ (always loaded) and _index.md (auto-maintained). Everything else is yours to shape.
- Best time to save: End of conversation — update everything at once without interrupting your flow
- Keep core small: Only put info needed in every conversation into
_core/. Everything else goes toprojects/ornotes/ - Works in any language: Trigger words work in English and Chinese (加载记忆, 记住这个, 记忆状态)
CLAUDE.md tells AI how to work on your project (coding style, tech stack, conventions). This skill tells AI who you are (identity, preferences, goals, knowledge). One manages the project, the other manages the person. They work together — not a replacement for either.
Skills may not activate in every conversation. For guaranteed memory loading, add this line to your CLAUDE.md:
When I say "load memory", read files from ~/.persistent-memory/_core/ and ~/.persistent-memory/_index.md
This ensures memory works even when the skill isn't triggered.
Works with any AI agent that supports the skills protocol. Tested with:
- Claude Code (Anthropic)
- Codex CLI (OpenAI)
- Antigravity (Google)
- Other agents on skills.sh
Since memory is stored as local files, all agents on the same machine share the same context automatically.
All memory is stored as local markdown files on your machine. Nothing is sent to external servers. Don't store passwords or API keys — memory files are plain text.
MIT