Agent-agnostic persistent memory for AI assistants. Your memories live in your own Google Drive as plain Markdown files.
Any AI that supports MCP can connect — Claude, GPT, Gemini, whatever comes next. Your AI changes, your memory stays.
Step 1: Authenticate with Google Drive (one-time)
npx cloud-memory-mcp@latest --installOpens browser → Google login → grant Drive access → done.
Step 2: Register in your AI client
# Claude Code
claude mcp add cloud-memory -s user -- npx -y cloud-memory-mcp
# Gemini CLI
gemini mcp add -s user cloud-memory npx -y cloud-memory-mcp
# Any MCP client: use stdio transport with command "npx -y cloud-memory-mcp"Restart your AI client. Memory tools are live.
npx cloud-memory-mcp --uninstallRemoves local tokens and MCP config. Your memories in Google Drive are untouched.
Agent-agnostic — Works with Claude, GPT, Gemini, or whatever AI comes next. Switch providers anytime, your memory follows.
Secure by default — Memories stored in YOUR Google Drive. Google's infrastructure, your account, your encryption. No third-party servers touching your data.
Permanent — Cloud storage persists across machines, OS reinstalls, and time. Not tied to any local filesystem or app.
Human-readable — Plain Markdown files you can open, edit, or delete anytime in Google Drive.
Virtually unlimited — Text-based memories take almost zero storage. Even thousands of memories barely scratch Google Drive's 15GB free tier.
| Existing Products | Problem |
|---|---|
| OpenAI Memory | ChatGPT-locked |
| Claude Memory | Anthropic-locked |
| mem0 | Their server, their rules |
| cloud-memory-mcp | Your data. Any AI. Forever. |
Use different Google accounts to naturally separate contexts:
- Personal Gmail → personal memories, preferences, life stuff
- Work Google account → work projects, meeting notes, professional context
Each account gets its own "Cloud Memory" folder. No config needed — just log in with the right account.
Default is Google Drive. Override with CLOUD_MEMORY_BACKEND env var:
| Backend | Value | Storage Location |
|---|---|---|
| Google Drive | gdrive (default) |
Cloud Memory/ folder in your Drive |
| GitHub | github |
Private repo my-ai-memory |
| Local | local |
~/.cloud-memory/memories.md |
Works out of the box. First run opens browser for Google login. Token saved locally, auto-refreshes.
Set GITHUB_TOKEN and CLOUD_MEMORY_BACKEND=github. Auto-creates a private repo my-ai-memory.
Each memory is an individual file in your Google Drive:
Cloud Memory/
profile.md ← stable identity info
2026-01-15_a1b2c3d4e5f6_works-at-acme-corp.md ← individual memory
2026-02-03_e5f6a7b8c9d0_prefers-dark-mode.md
2026-03-09_c9d0e1f2a3b4_started-learning-rust.md
Browse, edit, or delete memories directly in Google Drive. profile.md holds stable facts (name, job, preferences) — the AI updates it as it learns about you.
| Tool | Description |
|---|---|
add_memory |
Store a memory. Deduplicates automatically. |
search_memories |
Keyword search across all memories (uses Google Drive fullText search). |
get_recent_memories |
Load the most recent memories (default 128). |
read_block |
Read a memory block (profile, preferences, context). |
write_block |
Update a memory block. |
update_memory |
Edit a memory by ID. |
delete_memory |
Remove a memory by ID. |
The AI client (Claude, GPT, etc.) decides what to remember. The MCP server just stores and retrieves — no LLM calls, no API keys, no extra cost.
The MCP tools are available, but your AI won't use them unless instructed. Add a line to your agent's system prompt or instruction file:
Add the following to your agent's instruction file (CLAUDE.md, GEMINI.md, etc.):
- Use cloud-memory MCP tools for all memory. At conversation start: read_block("profile") + get_recent_memories. Store facts with add_memory, update blocks with write_block.
Running npx cloud-memory-mcp@latest --install will prompt to do this automatically for Claude Code (disables local auto-memory in ~/.claude/settings.json + adds the instruction to ~/.claude/CLAUDE.md).
- AI reads
profileblock + recent memories at conversation start - During conversation, AI stores new facts via
add_memoryor updates blocks - Each memory becomes an individual file in your Google Drive
- Next conversation — same context, any AI client, any machine
No vector DB. No embedding models. No infra. Just Markdown files in cloud storage you already own.
- Local cache layer for faster reads
- iCloud storage adapter
- Import from existing sources (CLAUDE.md, ChatGPT export)
Inspired by mem0 — the memory layer for AI agents. We adapted their insight that LLMs can manage their own memory lifecycle, but simplified the storage from vector DB to plain Markdown on user-owned cloud storage.
MIT