Every new Claude Code session starts from zero. You re-explain the architecture, the naming conventions, the decision you made last week, the bug you already fixed. The model is brilliant and amnesiac — and the amnesia is the expensive part.
This guide shows how to give Claude Code, Cursor, Claude Desktop, or any MCP-compatible client a persistent, cross-session memory in about a minute, using one config block and no infrastructure of your own.
Why not just use a bigger context window or a vector database?
Two common answers, two problems:
- A bigger context window doesn't persist. When the session ends, the context is gone, and pasting your whole history back in every time is slow and expensive.
- A vector database you host yourself means you're now running, tuning, and paying for an embeddings pipeline, a chunking strategy, and a retrieval service — before you've stored a single useful fact.
What you actually want is a memory service your agent talks to over a standard protocol, that survives every session and restart, and that you don't have to operate. That's what the Model Context Protocol (MCP) is for.
What is an MCP memory server?
MCP is the open standard that Claude, Claude Code, Cursor, and Cline already speak. An MCP memory server exposes a set of tools — store a fact, recall relevant context, bootstrap a session — that the model can call on its own. You add one server to your client config, and the model gains a long-term memory it manages itself.
Parametric Memory is MCP-native: 11 tools over streamable HTTP, OAuth2 + Bearer auth, and it works with any MCP client. No SDK to learn, no service to run.
Setup: from zero to remembering, in one config block
1. Get an instance and an API key. Sign up and pick a plan — your own isolated memory substrate provisions automatically. Then claim your API key from the dashboard; your personal MCP endpoint and key are shown there.
2. Add one MCP server to your client config. Paste this block, swapping in your endpoint and key from the dashboard:
{
"mcpServers": {
"parametric-memory": {
"type": "streamable-http",
"url": "https://<your-instance>.droplet-mcp.nz/mcp",
"headers": {
"Authorization": "Bearer <YOUR_API_KEY>"
}
}
}
}Where the config lives depends on your client:
- Claude Desktop / Cowork —
claude_desktop_config.json(full steps) - Claude Code — add it as an MCP server in your project or user config
- Cursor, Cline, and other MCP clients — same block, per-client location
3. Restart the client. Your agent now has memory. Ask it to remember something; start a new session tomorrow and ask it to recall — it's there.
What should your agent actually remember?
The highest-value things to store are the ones you keep re-explaining:
- Decisions and their rationale — "we chose Postgres over Dynamo because…"
- Conventions and rules — naming, formatting, "never touch the migrations directory"
- Corrections — when you fix the model's behaviour, store it so it stops recurring
- Project state — what shipped, what's in flight, what's blocked
Parametric Memory ranks and surfaces the relevant slice when a new session starts, so the model gets exactly what it needs without you pasting it back in.
The part vector memory can't do: proof
Most memory systems are probabilistic — you trust that what came back is what went in. Parametric Memory writes every fact into an RFC 6962 Merkle tree. If a single stored fact were altered, the root hash would change, so your agent (or you) can verify its own memory without trusting the server. You can check a signed snapshot yourself — no account, no API key, none of our code in the loop.
Cost
An isolated substrate starts at $5/mo (Starter) and $9/mo (Solo) on shared infrastructure, with $29/mo (Professional) and $79/mo (Team) on dedicated infrastructure. Every tier includes Merkle proofs, Markov prediction, knowledge-graph edges, and MCP-native access. See pricing.
Give it a try
If you use Claude Code, Cursor, or Claude Desktop every day, the amnesia is a tax you're paying on every session. One config block removes it — and unlike a memory you build yourself, this one is verifiable, predictive, and not your problem to run.