Every session with an AI coding assistant starts from zero. The constraints you explained yesterday, the decision you reached together last week, the context you carefully assembled — gone when the conversation ends. Worse, what you told Claude never reaches Cursor, and what Cursor learned never reaches ChatGPT. Each tool keeps its own partial picture, and none of it persists.
The Model Context Protocol (MCP) offers a clean fix: give every assistant read and write access to the same external store. This post explains the pattern and how to set it up with a notes vault as the memory.
Why assistants forget
Large language models have no memory between sessions by design — each conversation is a fresh context window. Vendors bolt on per-product memory features, but these are silos: tied to one tool, opaque to you, and invisible to every other assistant you use.
For anyone working across multiple AI tools — Claude Code in the terminal, Cursor in the editor, a chat assistant for research — this means repeating yourself constantly, and it means the most valuable byproduct of those sessions (decisions, findings, working notes) evaporates.
The pattern: external memory over MCP
MCP is an open protocol that lets AI clients call tools exposed by a server — search this, fetch that, create this. Claude Code, Claude Desktop, Cursor, Windsurf, and a growing list of clients speak it.
The memory pattern is simple: run your notes as an MCP server, connect each assistant to it, and the vault becomes shared memory.
Write path: "Create a note titled Daily Standup with my blockers for today." The assistant calls the server's create tool; the note now exists outside the conversation.
Read path: "Find all my notes tagged research and summarize the top 3." The assistant searches the vault and works from what it finds.
Because the store is external, three things fall out for free. Memory persists — next week's session can retrieve this week's decisions. Memory is shared — Claude and Cursor read the same notes, so context established in one tool is available in all of them. And memory is yours — it is a notes vault you can open, read, edit, and prune, not an opaque blob inside someone's product.
That last property is the important one. A memory you cannot inspect will accumulate stale and wrong context you cannot correct. A memory that is just notes gets the same curation your notes get.
Setting it up with Notez
Notez exposes workspaces over MCP, so any MCP-capable client can search and edit your notes in natural language. The setup:
Create or pick a workspace in Notez to act as the memory. A dedicated workspace keeps AI-written notes separate from personal ones, though sharing one vault is exactly the point for many workflows.
Generate an MCP key for that workspace from the integrations section of the web app. Keys are scoped to a single workspace and revocable at any time.
Point your clients at it — the integrations page gives you ready-to-paste config for Claude Code, Cursor, and Windsurf; any other MCP-aware client works the same way. Claude (web, desktop and mobile), ChatGPT and VS Code skip the key and connect by signing in, from the same page. (The Notez MCP server post walks through the per-client setup in detail.)
From then on, an assistant with the key has 31 tools: it searches and reads notes, creates and edits them, tags them, moves them between board columns, deletes them to the Bin and restores them. You interact in plain language; the assistant translates to tool calls. One design choice worth knowing: before an assistant changes a note's text, Notez saves the version it replaces. Version history keeps it for 30 days, and a note an assistant deletes waits in the Bin for the same 30 days.
What to actually store
Shared memory works best when you are deliberate about what goes in it. Categories that earn their place:
Decisions and their reasons. "We chose cursor pagination over offset because of table size" saves re-litigating it in every future session.
Project conventions. Naming rules, deploy quirks, the test command that actually works — the things you currently re-explain to every new conversation.
Findings from research sessions. When an assistant summarizes ten sources, have it write the summary to a note; the next session starts from the summary instead of the ten sources.
Open questions and blockers. Have the assistant write these down at the end of each session, and close them out yourself as they resolve.
What to keep out: secrets and credentials (memory is for context, not configuration), and raw transcripts — store conclusions, not conversations.
Retrieval is the other half
A memory that only accumulates becomes a junk drawer. Two things keep it useful.
First, meaning-based search. Notez indexes notes with embeddings, so an assistant asking for "notes about the caching approach" finds the relevant note even when its words do not match exactly. The same applies to you: the vault's Ask feature answers questions across everything stored, with sources.
Second, human curation. Because the memory is ordinary notes, you can read what your assistants have written, correct what is wrong, delete what is stale, and tag what matters. Skim the AI-written notes periodically, promote the durable ones, and prune the rest. Ten minutes a week keeps the memory trustworthy.
The compounding effect
The first week, this feels like a convenience — less repeating yourself. The compounding happens over months: a vault where your own captures and your assistants' notes accumulate side by side becomes a queryable record of how your projects actually evolved. You ask "what have we decided about the import pipeline?" and get an answer with sources, regardless of which tool — or which human — wrote the notes down.
Assistants will keep getting better, but they will keep starting conversations from zero. The memory that persists is the one you keep outside them.
Continue reading
Your notes in Claude and ChatGPT, without an API key — adding Notez to Claude, ChatGPT or VS Code with one URL and a sign-in, and taking access back.
Connect Notez to Claude, Cursor, and Windsurf with MCP — setting up an MCP key in Claude Code, Cursor or Windsurf, and the tools it gives an assistant.
Claude Code can now put images in your notes: file uploads over MCP — how an assistant with a shell puts a screenshot into a note without sending it through the chat.