The two are complementary rather than competing. Many setups keep a short CLAUDE.md that tells Claude which MCP memory tools to call and when, and keep the bulk of the history in the memory server.
How they differ
| CLAUDE.md and auto memory | MCP memory server | |
|---|---|---|
| Loading | Loaded at session start (auto memory: first 200 lines or 25KB of MEMORY.md) | Retrieved by tool call when relevant |
| Size | Anthropic suggests under about 200 lines per CLAUDE.md | Can hold far more than fits in context |
| Who writes it | You write CLAUDE.md; Claude writes auto memory | The assistant through tools, and you, depending on the server |
| Sharing | Project CLAUDE.md shared through version control; auto memory is per machine | Any client that connects to the same server |
| Other AI tools | Read by Claude Code (and AGENTS.md by several agents) | Any MCP-capable client the server supports |
| Setup | None | Install and configure a server |
When CLAUDE.md is enough
Claude Code loads project, user and local CLAUDE.md files at launch and treats them as context. Auto memory adds notes Claude writes from your corrections, stored per repository under ~/.claude/projects/<project>/memory/ and kept until you or Claude edit them. Imports (@path) let you split a long file, though imported content still costs context.
If your project rules fit in a couple of hundred lines, you work in one repository, and you mainly use Claude Code, this is free, versioned with your code and needs no extra software. Adding a memory server here mostly adds moving parts.
When an MCP memory server helps
Consider a memory server when the history has outgrown an always-loaded file, when you need to search source documents such as specs, PDFs or spreadsheets, when ChatGPT, Claude Desktop or Copilot must see the same project memory, or when several people need one record. Options include Basic Memory (Markdown notes, free local version), Mem0 OpenMemory, Supermemory (hosted), Pieces (captured work activity) and Central Brain.
Central Brain prepares selected project files locally, offers semantic and exact search with source path, page and chunk where available, and stores five note types plus a project brief as readable local files. Source-file tools are read-only; memory_checkpoint appends notes and central_brain_ping returns a RESUME block on connect. A practical pattern is a short CLAUDE.md line such as "At the start of a task, call central_brain_ping; checkpoint decisions when agreed," with the full record in the server.
Common questions
Does an MCP memory server replace CLAUDE.md?
Usually not. Keep project rules and build commands in CLAUDE.md, which always loads, and use the server for larger or shared memory that is retrieved when needed.
Is Claude Code auto memory shared with my team?
No. Auto memory is stored on your machine per repository. Share team rules through the project CLAUDE.md in version control.
Does connecting an MCP server send my files to Anthropic?
Claude receives the context the server returns to it. Review what each tool returns and your Claude plan settings; local storage and model processing are separate questions.