Guide
How to give Claude memory of your work
Claude forgets most of what you'd want it to know: the call this morning, the error in the browser, the decision you made in another chat. Here are the ways to fix that, from built-in features to MCP memory servers, and when each one makes sense.
What Claude knows by default
In a fresh conversation, Claude knows what's in its training data, what you type, and whatever you attach. That's it. It wasn't on your 10am call. It hasn't seen the stack trace in your browser. It doesn't know that you and a colleague agreed yesterday to drop the annual plan.
So most people do the same thing: they paste. The error, the doc, a paragraph of background. It works, but it's slow, you leave things out, and you do it again tomorrow. The fix is to give Claude a way to find that context itself.
Claude's built-in options
Memory
On plans that include it, Claude can remember things across chats: your role, your preferences, projects you've discussed. It's useful for “how I like things done”. It only knows what happened inside Claude, though, so it can't tell you what was said on a call or what you read in Notion.
Projects
A Project on claude.ai holds instructions and files that every chat in it can use. It's the simplest way to give Claude a stable set of documents: a spec, a style guide, a list of customers. The catch is that you have to keep those files current yourself.
Built-in memory is good at who you are. It doesn't know what happened today.
CLAUDE.md for Claude Code
Claude Code reads a file called CLAUDE.md at the start of each session. Put what doesn't change often in it: how to build and test, conventions, which folders hold what. Keep it short and specific. A good CLAUDE.md saves you repeating the same instructions every session.
What CLAUDE.md can't do is keep up with the work. Yesterday's decision, this morning's failing test, the thing a customer asked for on a call: by the time you've written those into a file, you could have pasted them.
MCP memory servers
The Model Context Protocol (MCP) lets Claude call tools. A memory server is a tool that stores and returns context. Claude decides when it needs something and asks. There are a few kinds:
- Notes you or Claude write. Knowledge-graph and Markdown servers where Claude saves facts as it goes. Good for decisions you want kept. They only contain what someone remembered to save.
- Vector stores. Servers that embed text and search it by meaning. Good for large document sets you load in.
- Activity capture. Servers that record what you actually do on your computer (screen, calls, files) so nothing has to be saved by hand. This is what Cachejay does.
The first two are good at what you put in them. Activity capture is the only one that knows about the call you didn't take notes on.
Which to use
| Option | Best for | Keeps itself current? | Knows about calls and screen? |
|---|---|---|---|
| Claude memory | Your preferences and background | Partly, from your chats | No |
| Projects | A fixed set of documents | No, you update the files | No |
| CLAUDE.md | Rules for a codebase | No, you edit it | No |
| Note-taking MCP servers | Facts Claude saves on purpose | Only what gets saved | No |
| Cachejay | What actually happened in your day | Yes | Yes |
Most people end up using a few of these at once. CLAUDE.md holds the rules, memory holds your preferences, and something like Cachejay covers what changes day to day.
Setting up Cachejay with Claude
- Install Cachejay and allow Accessibility (and the microphone, for calls).
- From the menu bar, choose Connect AI Apps → Claude Code or Claude Desktop.
- Ask Claude something only your day would know: “What did we decide on the last call about pricing?” or “Pick up where we left off on the billing service.”
For Claude on the web or your phone, add a custom connector with the URL https://cachejay.com/mcp and sign in with your Cachejay email. Claude can then search what you saw, said and decided, from anywhere.
More on what Claude can do with it: Cachejay for Claude.