engram-rs-mcp

engram-rs-mcp

MCP server for engram — persistent, brain-like memory for AI agents.

Works with Claude Code, Cursor, Windsurf, OpenClaw, and any MCP-compatible editor.

Quick Setup

{
  "mcpServers": {
    "engram": {
      "command": "npx",
      "args": ["-y", "engram-rs-mcp"],
      "env": {
        "ENGRAM_URL": "http://localhost:3917",
        "ENGRAM_API_KEY": "",
        "ENGRAM_NAMESPACE": ""
      }
    }
  }
}

Claude Code

claude mcp add -s user engram -- npx -y engram-rs-mcp

Cursor / Windsurf

Add the config above to .cursor/mcp.json or equivalent.

Environment Variables

Variable	Default	Description
ENGRAM_URL	http://localhost:3917	engram server URL
ENGRAM_API_KEY	—	Bearer token for auth
ENGRAM_NAMESPACE	—	Namespace for project isolation

Tools

Tool	Description
engram_store	Store a memory with optional importance, tags, kind
engram_recall	Hybrid search with budget, time filters, reranking
engram_resume	Session recovery — core + recent + topic index + triggers
engram_topic	Drill into topic clusters by ID from the knowledge index
engram_recent	List recent memories by time window
engram_search	Quick keyword search
engram_extract	LLM-powered text → structured memories
engram_consolidate	Run promotion/decay/merge cycle
engram_triggers	Pre-action safety recall
engram_stats	Layer counts and status
engram_health	Health check
engram_delete	Delete a memory by ID
engram_update	Update importance, tags, or layer
engram_trash	List soft-deleted memories
engram_restore	Restore from trash
engram_facts	Insert fact triples
engram_repair	Fix FTS index + backfill embeddings

What is engram?

Persistent memory organized along two dimensions — time (three-layer lifecycle: Buffer → Working → Core) and space (self-organizing topic tree). Automatic decay, promotion, dedup, LLM-powered quality gates, and topic clustering. Single binary, ~9 MB.

See the main repo for full documentation.

License

MIT