🧠 Engram

Portable, verifiable memory for AI agents — on Walrus. Tell your agent a decision, config, gotcha or fact about a teammate; it stores it on Walrus and recalls it across sessions, devices, models — and, in team scope, across your whole team. It stops re-asking for context you already gave it.

Built for the Walrus Memory hackathon (SuiHub Lagos, Aug 2026).

Two surfaces, one idea

• Engram MCP — add it to Claude Code / Cursor / Codex and your agent gets engram_remember / engram_recall / engram_tidy with personal vs team scope and typed capture rules. This is the reproducible artifact: real Walrus writes, verifiable blob ids.
• Engram web app — a chat agent that auto-captures durable facts, shows each one's Walrus blob id + inspect link, and has a 🔒 Personal ↔ 👥 Team toggle and a model switcher. Runs with zero keys via an offline mock.

The prompt

PROMPT.md is the copy-pasteable system prompt that drives it — the rules for what to remember and when to recall. Paste it into any agent wired to these tools.

Quick start

pnpm install

# The MCP (real Walrus memory in your agent):
pnpm login:memwal        # one-time browser sign-in
# → add mcp/engram-mcp.mjs to your client (see mcp/README.md), paste PROMPT.md

# The web app demo (no keys needed):
pnpm dev                 # http://localhost:5173

Proof it works

Memories written live to Walrus mainnet via engram_remember (aggregator returns HTTP 200):

• config → 2iD9wF…U2qdI
• decision → sXb0q3…TVgno

See SUBMISSION.md for the full write-up and VIDEO.md for the demo script.

Layout

PROMPT.md          the copy-pasteable memory protocol (the star deliverable)
mcp/engram-mcp.mjs the MCP server (scope + typed rules) + mcp/README.md
src/, api/         the web app (Vite/React) + its serverless API
scripts/           smoke tests and proof collectors