A CLAUDE.md or .cursorrules is a static file that goes stale. This is a directory the agent can search and write back into, so what was learned in one session is there at the start of the next one.
An MCP server over a `.context/` directory in your repository. That directory holds facts, decisions, regressions, preferences, session summaries and monthly changelog syncs; the server lets an MCP client search it and orient itself in it without you pasting anything.
- The context library searched, so an agent looks something up instead of rediscovering it
- Automatic orientation at the start of a session, from the index that loads every time
- Facts about architecture, database and deployment kept in files rather than in a chat
- Decisions recorded with their reasoning, so a settled argument is not reopened
- Regressions recorded, so a bug that was already fixed is not reintroduced
- Coding preferences that follow you across every repository, not just this one
Node and `npx`. `npx repomemory go` sets up the global profile, creates `.context/`, configures Claude Code and Cursor and runs the initial analysis — that path uses an API key. `npx repomemory setup cursor` needs no API key: it installs the MCP server into `~/.cursor/mcp.json`, adds the rules file, and lets Cursor's own model do the analysis when you run `/repomemory-analyze`.
One command — npx repomemory go