engram-walrus

Portable, verifiable memory for AI agents — on Walrus. An MCP server that gives any AI a memory it actually keeps. Store a fact in one AI, recall it in another. Personal and team scope, with typed capture rules.

▶ Live demo & docs: https://engram-pearl.vercel.app

Quick start (2 steps)

# 1. Sign in once (opens a browser — creates your own private Walrus account)
npx engram-walrus login

// 2. Add it to your AI client (Claude Code, Cursor, VS Code, Antigravity…)
{
  "mcpServers": {
    "engram": { "command": "npx", "args": ["-y", "engram-walrus"] }
  }
}

Then paste the Engram prompt into your agent's rules so it knows when to remember and recall. That's it — just talk, and it remembers.

Tools

| Tool | What it does | |---|---| | engram_remember(text, type, scope) | Store one typed durable fact — decision · config · gotcha · person · preference · plan. Also stores anything you ask it to remember/save/keep (a plan, note, snippet). | | engram_recall(query, scope) | Semantic recall — a few keywords or a whole sentence. | | engram_tidy(scope) | Report likely-duplicate memories to consolidate. |

scope is personal (private to you) or team (a shared pool). Each memory is Seal-encrypted and written to Walrus as a verifiable blob.

Team memory (shared, one URL)

For a shared team memory across people/devices, host the endpoint and point every teammate at one URL — see the team guide.

Notes

• Without login, it runs an in-memory demo store (no account needed) so you can test the tools. Run npx engram-walrus login for real, persistent Walrus memory.
• Node 20+. Credentials live in ~/.memwal/credentials.json.

MIT · built for the Walrus Memory hackathon.