Guide
What is an MCP memory server?
A plain-English explanation of the Model Context Protocol, what a memory server does for Claude, Cursor and Codex, the main kinds you'll find, and what to check before you install one.
MCP in one paragraph
The Model Context Protocol is an open standard for connecting AI apps to tools. An AI app (the client) connects to an MCP server, sees a list of tools with descriptions, and calls them when it decides they'll help. A tool might search your files, query a database or post to Slack. Anthropic introduced MCP in November 2024, and it's now supported by Claude, Cursor, Codex and a long list of other apps. Local servers usually run on your machine and talk to the app over standard input and output. Remote ones run on the internet over HTTP.
What a memory server does
A language model doesn't remember anything between conversations unless something gives it the history. A memory server is the something. It stores context and gives it back when the model asks.
The important word is asks. You don't paste the memory in. The model sees a tool like search_memory, reads its description, and decides to call it when you say “what did we decide about pricing?”. That's why the quality of a memory server depends on two things: what it has stored, and how well its tools are described so the model knows when to use them.
The main kinds
| Kind | How memory gets in | Good at | Weak at |
|---|---|---|---|
| Knowledge graph | The model saves entities and facts as you talk | Stable facts and relationships | Anything nobody saved |
| Notes / Markdown | You or the model write notes to files | Decisions you want kept, in a readable form | Staying complete |
| Vector store | Documents are embedded and indexed | Searching big document sets by meaning | Knowing what happened recently |
| Activity capture | Records what you do: screen text, calls, files | What actually happened, with nothing saved by hand | Capturing what was never on your computer |
The first three depend on someone deciding something is worth saving. Activity capture doesn't. It's closer to how your own memory works: you remember the meeting whether or not you took notes.
What to check before installing
- Where the data lives. On your machine, or on someone's server? Can you delete it?
- What leaves your machine. Some servers send content to a cloud service for processing. Know which, and when.
- What it never captures. For activity capture in particular: password managers, private windows, password fields.
- Prompt injection. Text from a web page you visited could contain instructions. A careful server marks recalled text as untrusted so the model treats it as data.
- Footprint. Screen capture can be heavy on CPU, memory and disk. Look for numbers.
- Which clients it supports, and how setup works for each.
An example: Cachejay
Cachejay is an activity-capture memory server for the Mac. It reads text through the macOS Accessibility API (no screenshots), transcribes calls on the Mac, keeps coding-agent sessions, and files it all by project. On the Mac it gives Claude Code, Claude Desktop, Cursor and Codex tools such as what_was_i_doing, project_brief, search_memory and todos. It also syncs to a private cloud space, with secrets masked, so ChatGPT and Claude on the web or a phone can search it through a remote server. Recalled text is marked untrusted, and you can keep everything on the Mac instead.
See the full list of tools in How it works, or how it fits with Claude in How to give Claude memory of your work.