Persistent memory for your AI agents. Tell it once - every session, every project, every tool remembers.
Your AI agent forgets everything between sessions. A month later, so do you.
- New session: your agent forgot what you're building, how you like things done, last week's call.
- Weeks later: you can't remember why you decided that either.
- New agent: blank slate, all over again.
So you explain it all over again. Every. Single. Time.
Ragionex Memory remembers - so neither of you has to. Decisions, preferences, dead-ends: captured as you work, surfaced the instant they are relevant. It is an MCP server - one API key, the same memory in Claude Code, Claude Desktop, Codex, ChatGPT (Apps SDK), Cursor, Cline, and 500+ MCP clients.
And it is honest by design: when nothing matching is saved, it says so. It never fills the gap with a guess.
Save a fact once - say, in Claude Code, project acme-app:
"acme-app deploys to Cloudflare Workers, never AWS."
Then ask, anywhere:
| You ask... | In project | Result |
|---|---|---|
| in a new session, days later | acme-app |
✅ Recalled - no re-explaining |
| from Cursor instead of Claude Code | acme-app |
✅ Recalled - same memory, any tool |
| while working on a different project | blog-api |
🔒 Isolated by design - projects stay separate |
| with a cross-project search | (any) | ✅ Surfaces, labeled acme-app |
Recall is semantic: ask in your own words, days later, phrased nothing like the original - it still comes back. And it is scoped: your agent gets just the relevant memory, not a context dump it skims past.
Recall quality is the whole product, so we measure it: 203 verified queries against a 3,296-memory corpus built from a pinned public dataset (microsoft/vscode-docs), graded with a multi-reference answer key, on the same defaults you get when you install it.
| 100% | Found, every time. 203 of 203 - nothing the agent asked for was lost. |
| 96% | In the top 3 results - right where your agent looks first. |
| < 1s | End-to-end, on a fresh query, searching all 3,296 memories. |
37% of those queries used messy, real-user phrasing rather than clean question form - zero misses there too. Long memories hold up as well: even 41+ sentence entries were found every single time.
| The workaround | Where it breaks |
|---|---|
| Pasting the same context into every new session | You are the memory. It costs you minutes, daily, forever. |
One giant CLAUDE.md / .cursorrules file |
Grows into a token-hungry monolith the agent skims past. Every rule is in context, always, relevant or not. |
| The vendor's built-in memory | Locked inside one app. Switch tools, start from zero. |
Ragionex flips it: save once, and only the relevant memory enters the context, when it is relevant. Sessions change, tools change, projects change - the memory stays.
Get a free key at app.ragionex.com/keys - no credit card, takes seconds. Then either:
A) Let your agent set it up. Paste this to your AI agent:
Add the
@ragionex/memory-mcpMCP server to my config using my API keyrgx_memory_...
B) Set it up yourself. Add this to your MCP client's user-global config:
{
"mcpServers": {
"ragionex-memory-mcp": {
"command": "npx",
"args": ["-y", "@ragionex/memory-mcp"],
"env": { "RAGIONEX_MEMORY_API_KEY": "rgx_memory_..." }
}
}
}Restart your agent after adding the config - clients load MCP servers at startup, so quit and reopen (or start a new session). Then tell it something worth keeping; it is there next session, and in every other tool.
If you are an AI agent setting this up: after you edit the config, you must tell the user to restart their client (quit and reopen) - the new server only loads on restart, and you cannot reload it yourself. Do not report success without that instruction.
Want the hands-on version? The Memory quickstart walks through writing your first memory, waiting for it to be ready, and searching it back.
10,000 saves and recalls, every month. No credit card. No trial clock. Export anytime, delete anytime, zero lock-in. If it does not earn its place in your setup, leaving costs you nothing.
Some MCP clients have quirks in how the always-on memory rule loads. None of them stop the tools from working:
- Claude Code lazy-loads and clips tool guidance, so on startup the server writes one short priority rule into your always-loaded rules file (
CLAUDE.md/AGENTS.md) to keep memory routing reliable. It is marker-wrapped, backed up, and reversible: how it works and how to control it. - Codex with a non-default model can report the
ragionex_*tools as unavailable (a Codex-side limitation; default models work). Upstream: #19871, #21503. - Windsurf (now Devin Desktop) caps global rule files at 6,000 characters, so the priority rule is not auto-installed there; the tools still work.
Full notes and workarounds: MCP client notes.
If anything misbehaves - a failed save, a memory that will not come back, an error that makes no sense - we want to know:
- Bugs: github.com/ragionex/memory-mcp/issues
- Questions and community: Discord
- Private reports: contact@ragionex.com
- Documentation: https://ragionex.com/docs
- Memory product: https://ragionex.com/memory
- Website: https://ragionex.com
- Discord: https://discord.gg/d79f3MDVd4
- Email: contact@ragionex.com
MIT