RMS Memory

Every AI coding agent starts from zero. Architectural decisions, known constraints, preferred patterns, debugging history — gone the moment you close the tab. You end up re-explaining the same stack to Cursor, then to Zed, then to Claude Code. Or you copy-paste a stale CLAUDE.md between tools and hope it's still accurate.

RMS Memory is a local-first memory layer for AI coding agents. A single centralized Markdown vault — structured for LLM consumption — that any MCP-compatible IDE reads from and writes to automatically. Context persists across sessions, tools, restarts, and team members.

Why existing solutions don't work

Most developers solve context loss with a CLAUDE.md or .cursorrules file. That works until you switch tools: Cursor doesn't read Zed's rules, Claude Code doesn't read Cursor's. You maintain copies, they drift, you paste manually.

The other approach — letting the AI summarize context into a new session — produces shallow, lossy reconstructions of decisions that took hours to reach.

Problem

• Every agent session starts from zero
• Rules files are tool-specific and drift out of sync
• Pasting context manually doesn't scale across 5+ active projects
• Semantic search over project knowledge doesn't exist

Solution

• One isolated vault per project, outside the repo
• All IDEs read from and write to the same vault via MCP
• Hybrid vector + full-text search finds relevant context even when keywords don't match
• Agents write decisions back proactively — vault grows as you work

RMS Memory works differently: instead of per-tool configuration files, a local MCP server maintains a structured Markdown vault and exposes it through five MCP tools. Any agent that supports MCP reads architectural decisions, constraints, and preferences from the same source. When an agent solves a problem worth remembering, it writes it back. The vault accumulates knowledge across sessions automatically.

Key Features

Hybrid retrieval — LanceDB combines vector similarity search with Tantivy full-text search. Queries never return empty because the exact keyword wasn't present. Reciprocal Rank Fusion merges results from vault and code corpora without incompatible raw distances.

Semantic code memory — optional Tree-sitter indexing for Rust, Go, TypeScript, JavaScript, Python, C, C++, Java, Ruby, Swift, and Vue. Agents search your codebase structurally — by function, struct, interface, or symbol — not just by filename or line content.

Global centralized vaults — project context lives outside your repo. No .mcp files, no per-repo pollution. A single ~/.rms-memory/registry.toml routes each project to its isolated vault. New IDEs connect automatically.

Multi-IDE auto-install — rms-memory install scans your system and wires itself into Cursor, Zed, Claude Code, OpenCode, VS Code, and Codex without manual JSON editing.

Audit metadata — every record automatically receives last_modified_by, timestamp, confidence, and source. Agents filter by reliability. Records without confidence are always included so pre-migration vaults stay usable.

Wiki isolation — the companion GUI can generate human-readable Wiki pages at <vault>/wiki/. The MCP server excludes this namespace entirely from indexing, search, watchers, and graph — so generated content never pollutes retrieval.

Production resiliency — atomic writes with rolling .bak backups, per-project index lock shared across IDE processes, panic-free database layer, path traversal protection, graceful shutdown, zombie process prevention. Ten-point resiliency checklist verified across concurrent multi-IDE sessions.

What the agent gets

• Persistent architectural context
• Structural code search
• Cross-IDE vault sync
• Confidence-filtered retrieval
• Proactive write-back

How it's implemented

• Rust
• LanceDB
• Tantivy FTS
• fastembed-rs (multilingual-e5-small)
• Tree-sitter
• Tauri 2 (GUI)
• WebGL (GUI graph)