Memstate AI - Git for Agent Memory
@memstate-ai
Git for agent memory. Memstate gives AI agents versioned, structured memory with automatic conflict detection, full version history, and up to 80% fewer tokens than vector-based alternatives. Extracts facts from text and markdown into structured keypaths, with 84.4% accuracy on t
Overview
What is Memstate AI - Git for Agent Memory?
Memstate AI - Git for Agent Memory is a hosted MCP server that provides versioned memory for AI agents. It stores facts, detects conflicts when information changes, and tracks decision history over time, using structured key-value storage instead of embedding-based semantic search.
How to use Memstate AI - Git for Agent Memory?
Get an API key from memstate.ai/dashboard, then configure your MCP client with the npx command npx -y @memstate/mcp and the MEMSTATE_API_KEY environment variable. No Docker or database is required. Supported clients include Claude Desktop, Claude Code, Cursor, Cline, Windsurf, Kilo Code, and Roo Code.
Key features of Memstate AI - Git for Agent Memory
- Structured, versioned key-value storage for agent memories
- Built-in conflict detection when new facts contradict old ones
- Deterministic recall — returns exact stored values, not approximations
- Constant O(1) token cost regardless of total memory size
- Full version history with soft-delete and tombstone preservation
- Hierarchical dot-notation keypaths for precise browsing
Use cases of Memstate AI - Git for Agent Memory
- Track architectural decisions and specifications across multiple agent sessions
- Detect and resolve contradictions when requirements or configurations change
- Maintain reliable cross-session context without dumping full history into prompts
- Audit how project knowledge evolved over time using full version chains
FAQ from Memstate AI - Git for Agent Memory
How does Memstate AI compare to RAG-based memory systems?
Memstate uses structured key-value storage with versioning rather than embedding-based semantic search. This gives deterministic recall, built-in conflict detection, and constant token cost, whereas RAG systems return approximate matches, cannot distinguish current vs outdated facts, and have O(n) token growth.
Do I need to run any infrastructure?
No. Memstate is a hosted SaaS — only an API key and the npx command are required. There is no Docker, database, or self-hosted server to maintain.
How are memories organized and retrieved?
Memories are stored in hierarchical dot-notation keypaths (e.g., project.myapp.database.schema). Keypaths are auto-prefixed with a project ID. The agent browses summaries first and drills into specific subtrees only when needed, keeping token usage constant.
What tools does the MCP server expose?
The server provides seven core tools: memstate_remember (for markdown summaries and decisions), memstate_set (for short scalar values), memstate_get, memstate_search, memstate_history, memstate_delete, and memstate_delete_project.
