Rms Memory MCP

What max-ramas/rms-memory-mcp MCP server does

max-ramas/rms-memory-mcp MCP server gives coding agents a persistent, project-scoped memory layer backed by local files and indexes. Each project can have an isolated vault stored outside the source repository, allowing several supported editors or agents to read and update the same context without copying notes into repository configuration files.

The server is designed for architectural decisions, requirements, user preferences, implementation notes, and session continuity. It also supports arbitrary scopes through the --scope option, so memory does not have to be tied only to a filesystem project.

How it works

Markdown content is chunked according to document structure, keeping headings associated with their lists and code blocks. Hybrid retrieval combines embedded vector search from LanceDB with Tantivy full-text search. Queries can target the vault, indexed code, or both; combined results use Reciprocal Rank Fusion rather than comparing raw scores from unrelated indexes.

Optional Tree-sitter indexing provides semantic code memory for Rust, Go, JavaScript, JSX, TypeScript, TSX, Python, C, C++, Java, Ruby, Swift, and Vue script blocks. Watch mode can reindex only changed paths, with a full scan fallback. The system can maintain up to four warm store-and-watcher pairs for concurrent project bindings.

Search uses bounded recall. rms_search can apply a character limit and minimum score, return an inject-or-abstain result, and fail closed when retrieval cannot meet its conditions. Short queries can use an FTS-preferred mode. Metadata such as confidence, source, timestamp, and the last modifying agent is attached to records for filtering and review.

Setup and configuration

The project publishes binaries for macOS Apple Silicon, Linux, and Windows through GitHub releases. macOS Apple Silicon and Linux users can install through the project’s Homebrew tap. A shell installer, PowerShell installer, source build, and an older crates.io installation path are also documented. The current README recommends Homebrew or release binaries over cargo install for newer workspace versions.

Build-from-source installation requires Rust 1.75 or newer. The resulting executable is named rms-memory. A dynamic installer can scan the system and configure supported IDEs, while rules-as-code patching can update files such as .cursorrules and .zed/assistant.md; these behaviors are opt-in where stated.

The optional RMS Memory GUI is a separate paid Tauri application. It uses the same vault but is not required to run the MCP server. GUI features such as Git synchronization, graph editing, and AI organization should not be treated as prerequisites for MCP operation.

Tools and capabilities

Named MCP operations include:

• rms_search for vault, code, or federated search.
• rms_code_search for code-oriented retrieval, including cross-project read-only federation.
• rms_write for durable Markdown writes with backup and atomic replacement behavior.
• rms_checkpoint_save, rms_checkpoint_done, rms_checkpoint_load, and rms_checkpoint_query for session checkpoints.
• rms_overview for project orientation and rms_system_instructions for self-bootstrap instructions.

Search requests can specify an explicit project and, where supported, a list of projects. Vault or all-corpus federation across projects requires the listed projects to permit cross_project_vault.

Limitations and notes

The server is local-first rather than a hosted synchronization service. Its supported project and editor behavior depends on configuration and the MCP client. The README names Cursor, Zed, Claude Code, OpenCode, and other MCP-compatible IDE workflows, but it does not establish compatibility with every MCP client.

Code indexing is optional and limited to the listed language grammars. The companion GUI is paid and may use a user-provided AI provider key for its AI-assisted features; that requirement does not apply to the core MCP server. Wiki files under the generated vault wiki directory are intentionally excluded from indexes, search, watchers, and graph data, and MCP writes reject those paths.