DEV Community

Ciro Cesar Maciel
Ciro Cesar Maciel

Posted on Originally published at blog.myagents.click

Giving Claude a memory that survives the conversation, through one MCP URL

You decided something with Claude yesterday: the project signs in with an e-mail code, no passwords. Today, in a new conversation, you explain it again. And again next week.

Memory for Claude means a decision made in one conversation can be found in the next one, without you repeating it. One way to get that is an MCP server that holds the memory, so every client you connect (Claude, Claude Code, Cursor) reads and writes the same place.

This is how it works with Brain, from Agents (by Ciro Cesar Maciel, who writes this).

One URL per space, no key

You create a space in the Brain dashboard, one per topic, and copy its address from the "Connection and settings" screen:

https://brain.worker.myagents.click/mcp/<space id>
Enter fullscreen mode Exit fullscreen mode

There is no key to paste. When the client calls the address without permission, it is told where to sign in, opens the browser, and you authorize that one space, read only or read and write. The token from that sign-in opens only that space; it does not work on the dashboard, and the dashboard's token does not work on the MCP address.

claude mcp add --transport http brain-work https://brain.worker.myagents.click/mcp/<space id>
Enter fullscreen mode Exit fullscreen mode

In claude.ai the same address goes in as a custom connector.

Six tools

Tool What it is for
search Finds what was already decided, by meaning and by exact word
resume Returns the work in progress, or the one for a given task
save Stores a decision, a preference, a fact or where something lives
revise Updates a memory that changed, keeping the previous version as history
forget Stops considering a memory, without deleting it
checkpoint Records the state of a task for the next session

Reading and writing are separate. Authorize read only, and Claude looks things up but never writes.

What a conversation looks like after that

You say, in passing, that the project signs in with an e-mail code. Following the rules the server hands over when it connects, Claude tends to save that as a decision, with the source (you) and the reason.

In a new conversation you ask for the sign-in screen. Before answering, Claude usually searches the space and, when it finds the decision, follows it. When the decision changes, it tends to revise instead of saving a second memory next to the old one. Leaving a task halfway, it records a checkpoint with what was done, what blocks and the next step, and resume brings it back next time.

What it saves by itself, measured

I measured it with real Claude Code sessions, in 14 natural-speech scenarios. Decisions said in passing, changed decisions, links, work to resume and requests to forget were saved without being asked. Passwords and keys were not saved, and memory was not touched in conversations that had nothing to do with it.

The weak spot is preferences said in passing. With Claude Code's own file memory on, 3 of 5 preferences went to Claude Code's files and 2 were not saved. With file memory off, or with one line in the client's instructions, all 5 went to Brain.

Checking what was saved

Everything Claude writes shows up under Memories in the dashboard, with author, source and date. The "To review" tab shows decisions that contradict each other, memories past their date and rules the agent wrote without a source. You decide what stands. Nothing there is deleted unless you type the space's name.

Honest notes

  • Brain stores what the agent or you tell it to save; it never receives the whole conversation.
  • The same space works in several clients, each after its own sign-in.
  • Free plan: 1 space, no limit on memories, no card and no expiry. Plans have 3, 9 or 99 spaces: https://myagents.click/en/products/brain#pricing

Top comments (0)