Walrus Ships MemWal — Decentralized, Verifiable Memory SDK Targeting AI Agents' Biggest Bottleneck

Walrus, the decentralized storage network on the Sui blockchain, launched MemWal on April 30 — an SDK that gives AI agents encrypted, verifiable, portable memory, addressing what enterprise customers and builders have called the largest unsolved bottleneck in agentic AI. MemWal saves memory records in encrypted form on the Walrus network, applies programmable access controls, and exposes cryptographic proofs that let any recipient verify a memory record has not been altered. Day-one integrations with the agent-orchestration frameworks OpenClaw and NemoClaw ship through plugins released the same week.

The bottleneck Walrus is pricing

Today's AI agents have a memory problem that compounds with every additional task: they either run inside a single chat session and forget everything when it ends, or they get bolted to centralized vector databases (Pinecone, Weaviate, Postgres + pgvector) that silo memory by vendor, lock customers into a specific cloud, and expose every memory record to the storage operator in plaintext. Neither pattern works for the use cases enterprises actually want to ship — agents that recall preferences across sessions, agents that hand work off to other agents, agents that operate inside regulated industries where data lineage and immutability are audit requirements. MemWal's pitch is a single primitive that solves all four: encryption (the operator can't read records), portability (memories migrate between agents and frameworks), verifiability (cryptographic proofs make tampering detectable), and availability (Walrus's storage guarantees survive any single node).

Why Walrus Foundation is making this bet now

The Walrus Foundation was incorporated in 2024 as an independent not-for-profit and raised $140 million in a private $WAL token sale in 2025. MemWal is the foundation's first major commercial product after building out the network's storage primitives, and the timing is deliberate. Agent infrastructure is the most valuable surface in AI right now: Stripe Link launched its agent wallet on April 30, OKX shipped its Agent Payments Protocol on April 29, Oobit issued Visa cards to agents the same week. None of those products solve memory; all of them assume a memory layer exists. AI Invest sized the addressable market at $70 billion over five years — a number that pencils out only if memory becomes the default storage layer for agentic systems, the way object stores became the default for cloud workloads. The OpenClaw and NemoClaw integrations matter because those are the two most-used open-source agent frameworks in the enterprise stack; getting MemWal as a default backend in either is more valuable than direct enterprise sales.

The skeptics' read

Three concerns. First, latency: decentralized storage has historically traded performance for guarantees, and AI agents need sub-second memory reads in tight loops. Walrus's published latency numbers are competitive with centralized object stores for read paths but lag on writes; whether the SDK's caching layer hides that gap is an empirical question. Second, standardization risk: every agent infrastructure provider is pitching a memory abstraction (Letta, Mem0, Zep, Cognee); MemWal is the first crypto-native entrant but not the first technical entrant, and developer mindshare is finite. Third, Sui dependency: MemWal runs on Walrus, which lives on Sui, which means an L1 outage or governance dispute on Sui hits MemWal directly — a trade-off enterprises with strong cloud-vendor preferences will scrutinize.

The competitive memory-layer landscape MemWal is entering

Agent memory is one of the densest pre-Series-B startup categories in AI infrastructure. The current field includes Letta (formerly MemGPT, $10M seed), Mem0 (Y Combinator W24, multi-tenant memory APIs), Zep (long-term memory for chat agents, Series A), and Cognee (knowledge-graph-based agent memory). All four operate on centralized SaaS infrastructure, which means memory records sit on the operator's servers, are readable by the operator's staff under most service agreements, and are not portable between agent frameworks without manual export.

MemWal's structural differentiation runs on three primitives the centralized incumbents cannot match: (1) operator-blind storage — Walrus's encryption-at-rest scheme means the storage operator literally cannot read records, which matters for regulated industries where vendor data access is itself a compliance issue; (2) cryptographic verifiability — proofs that detect tampering in transit or at rest, useful for legal and audit-trail use cases; (3) cross-framework portability — memories written by an OpenClaw agent can be read natively by a NemoClaw agent, or moved between organizations without re-vectorizing. The trade-off is latency overhead of decentralized storage versus a Postgres + pgvector setup running on a single AWS region; whether the SDK's client-side caching is good enough to close that gap on hot-read paths is the open empirical question.

What to Watch

Three signals over the next 90 days. Framework adoption: OpenClaw and NemoClaw integrations are day one; the next data point is whether LangChain, LlamaIndex, AutoGen or Anthropic's MCP ship MemWal as a supported backend. Enterprise reference customer: a named Fortune 500 deployment in regulated industries (finance, healthcare, government) is the credibility signal. Performance benchmarks: independently published latency and throughput numbers vs Pinecone, Weaviate and Mem0 will determine whether MemWal wins on guarantees alone or also on raw performance. Watch walrus.xyz/memwal and Sui ecosystem updates for the first concrete deployments.

Want every AI × Web3 signal the moment it breaks? Subscribe to the BlockAI News daily brief.