Walrus 🦭/acc

Imagine an AI agent kept a diary: Monday: "Started new project with user. Seems cool." Tuesday: "Started new project with user. Seems cool." Wednesday: "Started new project with user. Seems cool." This is how most production AI agents work today: no real memory between sessions, no awareness that you’ve talked before. @WalrusProtocol spent its first year on mainnet fixing this. Here's a quick TL;DR of their progress ↓ 1. What got built in 12 months → 510+ TB stored (passed Arweave's 385 TB inside a year) → 200+ projects building on it → 2nd-largest decentralized storage protocol by volume → Core research paper accepted at ACM CCS 2026, a rare academic stamp for any blockchain project → Zero downtime since day one The interesting stuff is what they built on top of it. Seal → Lets you encrypt data and control who can read it, enforced onchain. Made Walrus usable for healthcare records, financial data, and private AI datasets that can't sit on a public network. Quilt → Bundles small files together to make them cheap to store. Cut costs so much that it briefly reduced Walrus's own revenue. They shipped it anyway because users needed it. MemWal → The memory layer for AI agents. This is the one worth paying attention to. 2. MemWal It's an SDK that gives AI agents persistent, verifiable, portable memory that is encrypted by default and programmable through Sui smart contracts. Last week, they shipped a native plugin for NVIDIA’s NemoClaw and OpenClaw. Agents now auto-save and restore structured memories such as checkpoints, traces, history, and workflow state, as verifiable blobs on Walrus. What that means in practice: • An agent can save its work and pick up next session with cryptographic proof that nothing was tampered with. • Two agents in different companies can share memory securely. • You can switch between OpenAI and Anthropic models without losing context. • A regulator can verify exactly what data drove an autonomous decision. Now go back to the diary. Monday: “Started project with user. Background: prefers concise updates, ships on Fridays, three running threads.” Tuesday: “Continuing yesterday’s work. Made progress on thread 2.” Wednesday: “Sent Friday update. User happy.” That’s the difference. 3. Why this matters outside crypto Allium Labs → 65 TB of institutional blockchain data (the kind Visa, Stripe, Coinbase rely on) Tatum → 11 TB of historical Ethereum, Bitcoin, and BSC data, structured and programmable. OpenGradient → quietly crossed 4,000+ AI models with 4.5 TB+ stored on Walrus Team Liquid → 250 TB of esports archives Year one was about proving the storage actually works. Year two is about agents using it as their memory. Walrus is what gives them somewhere to remember. Disclosure: I'm a $WAL holder

x.com/i/article/2054…

Your AI agent has no memory. Every session starts from zero — re-reading docs, re-discovering your stack, re-learning what you already told it last week. We built OpenTusk so agents can own their memories. Launching on @WalrusProtocol on today.